Journal of Imaging doi: 10.3390/jimaging12080374

Authors: Yiping Li Ronald L. P. D. de Jong Franco Badaloni Gino M. Kuiper Romy C. van Jaarsveld Jelle P. Ruurda Marcel Breeuwer

Deep learning-based surgical vision systems achieve strong performance in semantic segmentation and phase recognition, but their black-box nature limits traceability in safety-critical clinical settings. Prototype-based networks offer an interpretable alternative by grounding predictions in learned visual exemplars, yet their suitability for surgical video understanding remains insufficiently characterized. We adapted a prototype-based architecture to two surgical datasets, laparoscopic cholecystectomy and robot-assisted minimally invasive esophagectomy (RAMIE), and benchmarked it against conventional baselines. We evaluated a segmentation-only setting, in which prototype size and capacity were ablated, and a multitask setting, in which three strategies for coupling prototype learning to semantic segmentation and surgical phase recognition were compared. Prototype-based models underperformed the conventional baselines across both tasks and datasets. In the segmentation-only setting, the selected prototype configurations achieved Dice scores of 72.23% on Cholecystectomy and 72.07% on RAMIE, compared with 74.35% and 74.02% for the corresponding conventional baselines, and showed weaker boundary agreement. In the multitask setting, the best prototype strategy recovered competitive segmentation performance but remained 6–12 F1 points below the conventional baseline for phase recognition. Qualitatively, prototype activation maps exposed intra-structure decompositions and contextual cues that are not directly available from black-box baselines. Prototype networks provide spatially traceable evidence for surgical scene understanding, but currently trade interpretability for reduced boundary precision and phase-recognition performance. These findings motivate future work on scene-level and temporally aware prototypes for explainable surgical AI.

Journal of Imaging doi: 10.3390/jimaging12080373

Authors: Tameto Naoi Jun Watanabe Keisuke Hamada Mitsuya Morita

Virtual reality-based subjective visual vertical (VR-SVV) has attracted attention as a potential solution to equipment-related limitations of conventional SVV testing. This study investigated the test–retest reliability and adequate trial number for stable assessment of vertical perception using VR-SVV in healthy adults. Participants performed 10 trials in a VR-SVV test and repeated the assessment after one week. SVV orientation and SVV variability were calculated. Test–retest reliability was evaluated using the intraclass correlation coefficient (ICC [1,2]), standard error of measurement (SEM), and minimal detectable change at the 95% confidence level (MDC95). The minimum number of trials required for stable assessment was examined by comparing results from fewer trials with those obtained from all 10 trials. SVV orientation and variability were −0.14° and 0.79°, respectively. ICC, SEM, and MDC95 were 0.62, 0.64°, and 1.76°, respectively. No adverse events occurred. SVV metrics derived from 6–9 trials differed by less than 10% from those obtained from 10 trials. VR-SVV is feasible and demonstrates moderate test–retest reliability. Six trials may be adequate for practical assessment of vertical perception.

Journal of Imaging doi: 10.3390/jimaging12080372

Authors: Zahoor M. Aydam Baidaa Mutasher Rashed Nidhal K. El Abbadi

Color texture classification is an important task in computer vision, with applications in medical imaging, industrial inspection, remote sensing, and material analysis. This paper presents a hybrid framework that integrates Principal Component Analysis (PCA), Local Binary Patterns (LBPs), Wavelet Scattering Transform, and the XGBoost classifier for color texture classification. The proposed pipeline first performs image pre-processing, including resizing and denoising, followed by channel-wise feature extraction using LBP and Wavelet Scattering Transform on the Red, Green, and Blue channels independently. Then, the obtained feature vectors were concatenated, and PCA was applied on the fused feature space for dimensionality reduction and redundancy elimination before proceeding to XGBoost classification. This method not only leverages complementary information of Chroma and texture information but also achieves reduced dimensionality and computational burden. The finally optimized features were input into the XGBoost classifier for color texture classification, which is good at fitting non-linear dependency and includes a regularization to generalize better. Our proposed framework was tested on three benchmark color texture datasets: KTH-TIPS, Outex_10, and VisTex. Experimental results have demonstrated that on these three datasets, the average performance reaches 98.0% accuracy, 0.981 precision, 0.981 recall, and 0.979 F1-score, respectively. It demonstrates that the two selected complementary feature extraction methods provide a compact yet effective representation for color texture classification on these datasets. It is expected that the proposed framework serves as an efficient combination of established methods and as a good competitive baseline for color texture analysis. Future works will consider applying it to larger color texture datasets for general verification, enhancing its computational efficiency and automating the parameter selection process.

Journal of Imaging doi: 10.3390/jimaging12080371

Authors: Yajnaseni Dash Vinayak Gupta Ajith Abraham Swati Chandna

In the original publication [...]

Journal of Imaging doi: 10.3390/jimaging12080370

Authors: Yuansa Wang Xueyao Gao Chunxiang Zhang Yongzeng Xue

With the rapid development of three-dimensional (3D) sensing technologies, multimodal 3D model classification has achieved significant progress. However, most existing methods are developed under closed and balanced assumptions, which limits their applicability to open long-tailed scenarios with scarce tail classes, ambiguous hard samples, and continuously emerging categories. In this work, we propose sample-enhanced category-space learning (SE-CSL) for open long-tailed multimodal 3D model classification. The proposed method first uses dual-branch modality encoders to extract point-cloud structural representations and multi-view semantic representations. Mamba is then introduced to model global dependencies across heterogeneous modalities and generate a unified global category representation. To improve the robustness of category representation, we design a category-space learning strategy that jointly integrates long-tailed learning, few-shot representation stabilization, and hard-sample enhancement. A long-tail balanced loss, a few-shot stabilization loss, and a hard-sample boundary loss are further developed to optimize intra-class compactness, inter-class separability, and boundary discrimination. To handle continually emerging classes, we introduce an incremental category-space expansion mechanism that distinguishes new classes, preserves old-class information, and supports unified classification of old and new categories. Extensive experiments on ModelNet40 and ShapeNet55 demonstrate the effectiveness and robustness of SE-CSL.

Journal of Imaging doi: 10.3390/jimaging12080369

Authors: Na Liu Fan Zhang Jiawei Wang Dan Zhang Jinliang Wu Xiaohui Li

Three-dimensional (3D) imaging systems, including depth cameras, LiDAR sensors, and multi-view scanning pipelines, often produce point clouds with noisy normals, outliers, sparse sampling, and non-uniform density, which can degrade downstream mesh reconstruction. Poisson surface reconstruction is lightweight and training-free, but its global implicit formulation is sensitive to unreliably oriented samples and fixed density-trimming thresholds. This paper presents RG-PSR, a reliability-guided enhancement framework for Poisson-family surface reconstruction from degraded 3D-imaging point clouds. RG-PSR estimates a deterministic per-point reliability score from local density regularity, spacing variation, and normal consistency, and propagates this score through conservative point filtering, reliability-guided normal refinement, adaptive density-reliability trimming, and structure-aware postprocessing. The main pipeline requires no manual labels, neural network training, or ground-truth meshes at inference time. Experiments on three groups of object meshes under five deterministic degradation types show that RG-PSR improves Poisson-family reconstruction under degraded inputs. Compared with fixed density-trimmed Poisson reconstruction, RG-PSR reduces the overall Chamfer-L1 from 0.0218 to 0.0172, improves F0.01 from 0.6618 to 0.6836, and reduces Artifact0.02 from 0.3090 to 0.2632. In the broader classical comparison, local triangulation methods achieve stronger point-wise accuracy, while RG-PSR yields the fewest connected components and the highest largest-component ratio. These results position RG-PSR as a practical reliability layer for coherent Poisson-family reconstruction rather than a universal replacement for all surface-reconstruction methods.

Journal of Imaging doi: 10.3390/jimaging12080368

Authors: Simona Moldovanu Adina Cocu Diana Stefanescu Cătălin Anghel

In recent years, there has been increased attention on classifying histopathological images through hybrid decision-level fusion, and the challenge of exploring data fusion to improve classification accuracy in colon cancer has become significant. This study introduces new elements by incorporating various deep learning (DL) architectures, including EfficientNetB0, DenseNet121, ResNet101V2, NASNetMobile, MobileNetV2, and VGG16 Convolutional Neural Networks (CNNs), as well as Random Forest (RF) and Histogram Gradient Boosting (HGB) Machine Learning (ML) algorithms, along with the original dataset. The proposed hybrid decision-level fusion approach analyzes the LC25000 dataset’s colon histopathological images and handcraft features (HFs) to improve predictive performance. The HFs such as entropy, the Gini index, and the radius of gyration from adenocarcinoma and benign colon tissue (CC) were extracted. The prediction of the proposed models leveraging late fusion was conducted by classifying both deep and HFs. During the experiments, it was demonstrated that the combination of ResNet101V2 with RF classifier and all HFs yielded greater accuracy and consistent performance. The achieved performance metrics include accuracy of 94.1%, F1-score of 94%, Matthews Correlation Coefficient (MCC) of 88.3%, and an area under the curve (AUC) of 0.979. To explain and interpret the decisions made by the DL models, the explainable methods SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM) were utilized.

Journal of Imaging doi: 10.3390/jimaging12080367

Authors: Zijun Wang Hongcheng Li

Semantic communication provides a task-oriented alternative to conventional bit-level transmission. For wireless image transmission, existing systems mainly optimize image reconstruction, while downstream classification is often handled by a separate model. This paper proposes Mission-ADWITT, a multi-task semantic communication framework that extends the ADWITT backbone with a classification branch and jointly optimizes image reconstruction and classification. To improve joint training, the reconstruction backbone is initialized from a pretrained ADWITT-CIFAR10 model, while the classification branch is randomly initialized and fine-tuned together with the backbone. Experiments are conducted on CIFAR-10 over AWGN channels at SNR values of 0, 5, 10, 15, and 20 dB. Compared with Mission-ADWITT without ADWITT initialization, the baseline-initialized model improves average classification accuracy from 75.156% to 82.988%, PSNR from 27.933 dB to 30.828 dB, and MS-SSIM from 0.9637 to 0.9804. It also achieves classification accuracy comparable to the separate ADWITT + ResNet18 baseline, although reconstruction quality remains lower. These results highlight the importance of reconstruction-pretrained initialization and the trade-off between visual fidelity and semantic performance.

Journal of Imaging doi: 10.3390/jimaging12080366

Authors: Jonathan Lee Christian Lee Eric D. Cyphers Casey Bonzell Simon Kim Michael Repajic Reza Assadsangabi Thomas G. Clifford Vinay Duddalwar Bryce D. Beutler

Whole-body magnetic resonance imaging (WB-MRI) has become established for selected clinical indications and is increasingly studied across many fields including but not limited to oncologic, musculoskeletal, pediatric, and computational imaging applications. This study characterizes recent trends and the evolution of WB-MRI research using a bibliometric analysis reported using the PRISMA 2020 guidelines and identifies dominant clinical, quantitative, and artificial intelligence (AI) themes shaping the field. Publications were identified from Scopus, Web of Science, and PubMed using WB-MRI and its variant search terms and filtered to articles and reviews from 2015 through to 2025. Bibliometric analyses summarized publication output, contributors, citations, and keywords, and used exploratory keyword rules to classify major clinical and technical themes. Of 1511 included WB-MRI publications, annual output increased from 124 in 2015 to 180 in 2025. The journals with the highest publication counts were European Radiology, PLoS ONE, and the European Journal of Radiology. Parsed author affiliations most frequently represented the United States, Germany, and the United Kingdom. Research centered on oncologic applications, particularly myeloma and prostate cancer, alongside musculoskeletal disease, diffusion-weighted imaging, and AI-based segmentation. AI-related publications increased from 5 in 2015 to 20 in 2025. Broader clinical use of WB-MRI will require standardized acquisition, reproducible quantitative measures, and multicenter validation.

Journal of Imaging doi: 10.3390/jimaging12080365

Authors: Liang Lyu Hao Sun Jiaqing Liu Fang Wang Lanfen Lin Yen-Wei Chen

Synthesizing multi-phase contrast-enhanced CT images from non-contrast CT (NCCT) may provide complementary phase-specific cues for preliminary assessment. After rigorous clinical validation, such images could provide clinical decision support or aid diagnostic triage by identifying cases that warrant further acquired contrast-enhanced CT (CECT) work-up; they are not intended to replace acquired CECT. Arterial-phase (ART) and portal-venous-phase (PV) images share anatomical structures but exhibit distinct enhancement patterns and intensity distributions, which makes simultaneous multi-phase synthesis challenging for conventional single-decoder models. We propose a task-prompt-guided dual-decoder framework that combines a shared Swin Transformer encoder, two phase-specific decoders, learnable task prompts initialized from Qwen3-8B semantic representations, two phase-specific adversarial discriminators, and an independent ART/PV domain classifier. The shared encoder extracts phase-invariant anatomical features, whereas the two decoders independently model ART- and PV-specific enhancement. The Qwen3-derived prompt vectors provide phase-aware initialization and subsequently adapt through prompt–feature interaction at the bottleneck. Experiments on a single-center dataset of 86 patients show improved whole-image and lesion-focused PSNR, SSIM, MSE, and PCC relative to Pix2pix and MedGAN. Controlled ablation experiments indicate the benefits of Qwen3-based prompt initialization, subsequent prompt adaptation, and correct prompt–decoder correspondence; a reduced baseline additionally assesses the combined removal of LLTP and ART/PV domain classification. In a downstream slice-level four-class focal liver lesion classification experiment, synthetic multiphase input improved accuracy from 71.65% with NCCT alone to 85.04%, compared with 91.34% for acquired multiphase CT. These findings provide a proof of concept for LLM-guided multi-phase CT synthesis, although external validation, clinically oriented safety assessment, and reader studies remain necessary before clinical use for decision support or diagnostic triage.

Journal of Imaging doi: 10.3390/jimaging12080363

Authors: Jiajing Zhou Haolin Wang Ikuma Nakagawa Shun Tanimura Yuhei Shibata Ken Nagahata Tamotsu Kamishima

Rheumatoid arthritis (RA) frequently involves the wrist joints, where early inflammation may induce subtle changes in carpal spatial architecture. This study investigated whether baseline (BL) ultrasound inflammation is associated with longitudinal changes in AI-derived compactness metrics and whether these metrics provide information complementary to conventional wrist radiographic scores. This study included 159 patients with early RA who underwent BL and follow-up (FU) bilateral wrist radiography and BL wrist ultrasound. Four AI-derived centroid-based compactness metrics—root-mean-square-radius (RMSR), mean pairwise distance (MPD), median radius (R50), and 90th-percentile-radius (R90)—were calculated, and their annualized changes were evaluated in relation to conventional wrist SvdH scores and BL ultrasound inflammation, including gray-scale (GS), power Doppler (PD), GS+PD, and vascularity percentage (VS%). Associations were assessed using correlation and subgroup analyses. AI-derived compactness metrics at BL and FU, as well as their annualized changes, were not significantly associated with conventional wrist radiographic scores. In contrast, BL ultrasound inflammation was significantly associated with subsequent increases in compactness metrics. BL GS, PD, and GS+PD scores showed significant positive correlations with annualized changes in RMSR, MPD, R50, and R90 (r = 0.18–0.25, p < 0.05). Subgroup analyses similarly demonstrated greater increases in compactness metrics in patients with positive GS or PD activity. Annualized changes in AI-derived compactness metrics were associated with baseline ultrasound inflammation but not with changes in conventional wrist radiographic scores. These findings suggest that AI-derived compactness metrics may provide complementary quantitative information on global carpal spatial architecture in early RA.

Journal of Imaging doi: 10.3390/jimaging12080364

Authors: Quang Tuan Ho Ngoc Ha Bui Thuy Duong Tran Quang Huy Khuat Ngoc Toan Tran Xuan Chung Le Huu Quyet Nguyen Tat Thang Nguyen Van Thai Nguyen Dinh Thuy Mai Quang Duy To Dinh Chau Nguyen Nguyen Huong Giang Trinh Van Chinh Cao Tien Hung Bui Thu Trang Vu Khac Nam Vo Hai Quan Ho

Accurate segmentation of non-small cell lung cancer (NSCLC) on positron emission tomography/computed tomography (PET/CT) is an essential prerequisite for automated metabolic tumor volume (MTV) quantification and staging. Although deep learning models achieve high performance on large-scale datasets, their generalization across different clinical domains is limited by variations in imaging protocols and patient demographics. This study aims to evaluate several deep learning architectures and investigate a transfer learning strategy to mitigate domain shift. Three architectures, including ResNet-backbone 3D U-Net, nnU-Net v2, and Swin UNETR, were benchmarked from scratch and compared with a fine-tuned nnU-Net initialized with AutoPET II weights. Results on the internal dataset showed that the fine-tuned nnU-Net achieved a Dice similarity coefficient (DSC) of 83.4 ± 6.5%, a 95% Hausdorff distance (HD95) of 5.1 ± 3.6 mm, and a precision of 89.6 ± 8.2%. Compared to the nnU-Net v2, the fine-tuned nnU-Net improved the absolute DSC by 6.8% while reducing local training time by 37.5% by bypassing the initial feature-learning phase. The fine-tuned nnU-Net model also demonstrated a high correlation between the MTV and the ground truth (Pearson r = 0.96, p < 0.001), indicating its potential as a reliable automated approach for quantitative MTV extraction and NSCLC prognostic-related analysis.

Journal of Imaging doi: 10.3390/jimaging12080362

Authors: Qing Chen Chao Wei Bingmeng Zhu Qiang Yan Shengbing Chen Dongmei Zhou

Self-supervised monocular depth estimation is challenging at night because adverse illumination degrades the visual cues required for correspondence estimation and depth inference. Although appearance compensation and domain transfer can mitigate nighttime visual variations, reliable depth recovery under such conditions still depends on exploiting incomplete local structural cues and uncertain scene context. To address this challenge, we propose the Dual-Path Interaction and Dual-Attention Fusion Network (DIDAF-Depth), which improves nighttime depth recovery through coordinated convolutional neural network (CNN)–Transformer interaction, attentional feature fusion, and structure-preserving reconstruction. Specifically, we design the Transformer-CNN Vertical Interaction Fusion (TC-VIF) encoder to perform bidirectional cross-layer exchange, allowing local structural cues and global scene context to complement and progressively refine one another during feature extraction. We further develop the Dual-Coupled Attentional Fusion Module (DCAFM) to model spatial and channel interdependencies and selectively integrate complementary local and global information into a unified representation for depth decoding. Building on DCAFM’s unified representation, we construct the Edge-aware Densely Cascaded Multi-scale Network (EDCMN) to propagate features across scales, reinforce weak boundaries during upsampling, and preserve structural continuity in predicted depth maps. Experiments on the nighttime subsets of the Oxford RobotCar and nuScenes datasets indicate that DIDAF-Depth provides strong and consistent performance under the adopted evaluation protocols, supporting the effectiveness of the proposed framework.

Journal of Imaging doi: 10.3390/jimaging12080361

Authors: Zainab Shanta Swayedi Pedram Salehpour Hadi Aghdasi

Accurate mammographic mass segmentation is crucial for computer-aided breast cancer diagnosis, but remains challenging because masses may be small, low-contrast, irregularly shaped, and partially obscured by dense fibroglandular tissue. Although recent methods have improved contextual representation and multi-scale feature extraction, reliable segmentation still requires better integration of lesion-scale representation, contour-derived auxiliary supervision, and local probability-map refinement. This paper proposes MB-SwinRefiner, a novel two-stage framework consisting of a Multi-scale Boundary-aware SwinUNet (MB-SwinUNet) stage and a probability-guided residual refinement stage. The first stage, MB-SwinUNet, generates an initial mass probability map using hierarchical Swin encoding, multi-scale decoder fusion, and training-only auxiliary boundary supervision. The second stage refines this prediction through probability-guided residual correction by using the Stage 1 probability map as a prior and learning logit-space corrections. Experiments were conducted on a case-level five-fold cross-validation of the INbreast dataset under both a resized benchmark setting and a native sliding-window full-mammogram setting. Performance was evaluated using Dice, intersection over union, sensitivity, and specificity. In the resized benchmark setting, MB-SwinRefiner achieved Dice of 86.19%, IoU of 84.33%, sensitivity of 89.34%, and specificity of 99.98%. In the native sliding-window setting, MB-SwinRefiner improved mean Dice from 76.64% to 78.06% and mean IoU from 74.38% to 76.11%. These results suggest that combining Swin-based contextual encoding, multi-scale decoder fusion, contour-derived training supervision, and probability-guided residual refinement can improve region-based mammographic mass segmentation performance.

Journal of Imaging doi: 10.3390/jimaging12080360

Authors: Sasan Karamizadeh Saman Shojae Chaeikar Mazdak Zamani

Facial expression recognition technology is vital for security, verification, and personalization, but it faces challenges due to variations in scale, illumination, occlusion, and facial expressions. This paper presents a hybrid architecture that combines Vision Transformers (ViTs) to capture global context with EfficientNet-B3 for multi-scale feature extraction. Unlike simple concatenation, our approach projects the ViT’s [CLS] token and the EfficientNet’s global pooling features into a shared 512-dimensional space before merging, enabling better alignment of global and local features. When tested on the FERPlus dataset, it reaches an accuracy of 94.4 ± 0.3%, surpassing several recent methods, notably existing transformer- and CNN-based methods. Ablation studies show each component’s contribution, with the full model outperforming the no-fusion version by 2.6%. With around 98 million parameters and an inference time of ~23 ms per image, it balances efficiency and high performance, suitable for real-time use on suitable hardware. Evaluation via confusion matrix, t-SNE visualization, and comparisons with recent techniques such as HLA-ViT (90.13%), AU-ViT (90.15%), and CCFER (91.24%) demonstrates its robustness and discriminative feature learning. This work highlights the promise of hybrid deep learning architectures in tackling real-world facial expression recognition challenges.

Journal of Imaging doi: 10.3390/jimaging12080359

Authors: Cynthia Martínez-Ramos Citlalli Gaona-Tiburcio Erick Maldonado-Bandala Demetrio Nieves-Mendoza Laura Landa-Ruíz Maria Lara-Banda Francisco Estupinan-Lopez Miguel Angel Baltazar-Zamora Jesús Manuel Jáquez-Muñoz Jose Cabral-Miramontes Facundo Almeraya-Calderón

Corrosion remains one of the principal causes of degradation in metallic structures across a wide range of industrial sectors. Although visual inspection is routinely employed for preliminary corrosion assessment, its effectiveness depends heavily on operator experience and subjective interpretation. This work introduces CorrQuant, a web-based computer vision platform designed to transform qualitative corrosion images into quantitative measurements of corrosion extent and morphology. The proposed methodology processes images acquired with conventional mobile devices and integrates geometric calibration using a reference coin, perspective correction, adaptive image enhancement through Contrast Limited Adaptive Histogram Equalization (CLAHE), multi-descriptor feature extraction, and consensus-based corrosion segmentation. The detected corrosion regions are subsequently quantified to determine corrosion area, surface coverage, spatial distribution, morphological descriptors, and corrosion intensity maps. The methodology was verified using an aluminum specimen with a known corrosion area of 143 mm2 under both controlled illumination and optical stress-test conditions. Under standard acquisition conditions, corrosion-area estimation accuracies exceeding 90% were achieved. Additional evaluations under red illumination, fisheye, blur, and kaleidoscope distortions demonstrated that the proposed framework is considerably more sensitive to degradation of local image information than to variations in illumination spectrum. These results demonstrate the robustness of the proposed multi-descriptor voting strategy while defining the operational limits of the platform under challenging image acquisition conditions.

Journal of Imaging doi: 10.3390/jimaging12080358

Authors: Syndar Satbayev Didar Yedilkhan Aruzhan Shoman Azamat Serek Mohammad Shadab Khan

Urban tree mapping is necessary in environmental sustainability and climate change mitigation, and it depends heavily on the individual tree recognition and canopy segmentation to analyze city green cover. This systematic review discusses recent developments in the 2014–2026 mapping of these trees with the use of UAVs and high-resolution satellite imagery. Our preliminary selection of 4148 records reduced to 101 eligible publications following a systematic screening and synthesis of the records, assessed the efficiency of deep learning models such as Convolutional Neural Networks and Vision Transformers in processing various source images. We also compare object detection and semantic segmentation to see which one is more competent to deal with typical urban challenges, including overlapped canopies and building shadows. According to the reviewed studies, UAV-based models generally achieve higher spatial accuracy than satellite-based approaches for individual tree detection and crown delineation, with reported average Intersection over Union (IoU) values of approximately 70–75%, whereas satellite imagery provides superior spatial coverage for large-scale urban forest monitoring. Lastly, we present a research roadmap to address the existing weaknesses such as geographic bias, which propels the research direction towards multimodal data fusion and Foundation Models to sustain consistent, large-scale urban forest monitoring.

Journal of Imaging doi: 10.3390/jimaging12080357

Authors: Aigerim Yerimbetova Ulmeken Berzhanova Marek Milosz Bakzhan Sakenov Elmira Daiyrbayeva Lyailya Cherikbayeva

With the rapid advancement of sensor technologies, automated sign language recognition (SLR) has emerged as a critical enabler of inclusive communication systems for individuals with hearing and speech impairments. Although substantial research effort has been directed toward this domain, existing reviews lack a structured comparison of sensing modalities and do not systematically address the challenges of low-resource sign languages. This paper presents a comprehensive systematic review of sensor-based and multimodal SLR systems, covering 76 publications from 2021 to 2026 selected through a PRISMA 2020 protocol. We propose an original four-category taxonomy encompassing wearable sensor-based, contactless non-visual, vision-based, and multimodal systems, and provide a three-category methodological classification distinguishing conventional, machine learning, and deep learning approaches. The comparative analysis reveals that, despite notable progress, critical challenges persist: the absence of standardized datasets, limited cross-user generalization, insufficient multimodal fusion strategies, and inadequate representation of low-resource sign languages, including Kazakh Sign Language (KSL). The findings of this review establish a structured foundation for future research aimed at developing robust, scalable, and computationally efficient SLR systems.

Journal of Imaging doi: 10.3390/jimaging12080356

Authors: Aikaterini I. Nikolaou Nikitas Chatzigiannis Maria Alexandra Kefala Eleni Papaioannou Maria Baltogianni Lida-Eleni Giaprou Vasileios Giapros

Neonatal echocardiography remains an essential imaging modality for the assessment of congenital and hemodynamic cardiovascular abnormalities in critically ill neonates. Recent advances in quantitative ultrasound biomarkers, deformation imaging, volumetric reconstruction, and artificial intelligence (AI)-assisted analysis have substantially expanded the diagnostic capabilities of neonatal cardiovascular ultrasound imaging. This narrative review critically examines current developments in quantitative echocardiographic imaging, advanced volumetric methodologies, and AI-assisted cardiovascular ultrasound analysis, with emphasis on neonatal intensive care applications. A literature search was conducted using PubMed, Google Scholar, Scopus and Embase focusing on neonatal echocardiography, spatiotemporal image correlation (STIC), speckle-tracking echocardiography, artificial intelligence, machine learning, and quantitative cardiovascular imaging. Recent studies suggest that advanced methodologies, including speckle-tracking echocardiography, STIC-based reconstruction, automated segmentation algorithms, and deep learning frameworks, may improve image standardization, automated quantification, and congenital heart disease detection. However, important challenges persist, including operator dependency, dataset heterogeneity, limited external validation, cross-platform variability, and incomplete integration into routine neonatal intensive care workflows. Most AI-assisted echocardiographic systems remain investigational and require further prospective multicenter validation before widespread clinical implementation can be achieved.

Journal of Imaging doi: 10.3390/jimaging12080355

Authors: Chaowei Dong Zhaodong Chen Rongwei Fan Zhiwei Dong Deying Chen Pengfei Hao Lansong Cao

Wide-area building-structure mapping from long-range airborne LiDAR requires image products that can represent building footprints, roof-height variations, and structural discontinuities with low computational latency. Airborne streak-tube imaging LiDAR (ASTIL) records a spatial–temporal echo image for each laser pulse, where the detector row corresponds to the fan-beam spatial angle, the detector column encodes echo arrival time, and the frame sequence represents the scanning process. This row–column–frame topology makes it possible to generate image-domain structural products directly from raw streak-tube echo sequences. In this paper, a relative-height image generation method is proposed for long-range ASTIL. The method constructs slant-range matrices from raw echo images, suppresses row-wise ground-related range trends, maps the residuals into relative-height values, and generates scan-geometry-calibrated swath-level relative-height images using lightweight calibration rather than rigorous point-wise POS/IMU trajectory reconstruction. Airborne experiments at 2 km, 3 km, and 6 km flight heights show that the proposed workflow can generate relative-height images with spatial sampling intervals of 0.30 m, 0.45 m, and 0.90 m, respectively, within a 0.5 s acquisition window. The generated cropped image products occupy less than 0.3% of the raw streak-image sequence volume, reflecting a compact image-domain representation for rapid preliminary mapping rather than lossless data compression. Building-scale comparisons with UAV LiDAR reference data indicate that the generated images preserve the main building footprints, boundary orientations, and roof-height discontinuities. For nine flat-roof targets, the mean absolute roof-to-ground height errors range from 0.24 m to 0.30 m across the three flight heights. These results suggest that ASTIL relative-height imaging can provide an efficient image-domain representation for wide-area building-structure mapping under long-range airborne observation conditions.

Journal of Imaging doi: 10.3390/jimaging12080354

Authors: Ivan Melegatti Fernigrini Bensheng Yun

Animal re-identification (Re-ID) asks whether two images show the same individual, a recognition task that fits naturally into applications such as reuniting lost pets with their owners. Existing methods report strong scores, but typically under a single seed, one mask granularity, and no query-time corruption analysis, leaving open whether the gains survive deployment. We propose a hierarchical framework decoupling localisation (a YOLOv8 soft-crop) from identity embedding: a dual-stream network fusing a frozen CLIP ViT-B/16 (learned projection) with a fine-tuned ViT-Base carrying L2-norm part attention, trained under ArcFace. On a combined cat+dog open-set benchmark of 173 identities, it attains Rank-1 0.9742/mAP 0.8597 over three seeds, surpassing a ViT-only ablation by +2.39 Rank-1 and +2.07 mAP. Open-set verification shows all configurations converge near 68% true acceptance at the strictest false-acceptance rate. A background-bias evaluation brackets the embedding’s background reliance between a bounding-box lower bound and a SAM-silhouette upper bound; a manual audit retains the reliable soft crop. A nine-corruption audit identifies down-sampling and motion blur as dominant. On PetFace, the architecture retrieves across 14,716 unseen identities and remains viable in a few-shot regime. A Descriptor Vector Exchange (DVE) extension is Pareto-dominated, traced to the ViT’s coarse feature map and architectural redundancy.

Journal of Imaging doi: 10.3390/jimaging12080353

Authors: Lizhuo Liu Jiawei Niu Lingxia Mu

Infrared target recognition plays an important role in many real-world applications, but its performance is often constrained by the scarcity of annotated infrared data. To alleviate this issue, optical-to-infrared image translation has been widely explored as a data augmentation strategy by leveraging the abundance of optical images. However, existing approaches typically overlook the intrinsically multimodal nature of optical-to-infrared mapping, leading to insufficient diversity in the synthesized infrared images. Moreover, the lack of effective constraints to preserve semantic fidelity further hampers the practical utility of generated samples for recognition tasks. In this paper, we propose a multimodal style translation framework for infrared target recognition. The proposed framework is built upon a style-semantic disentanglement architecture, which decouples domain-general semantic structures from domain-specific style statistics, thereby enabling flexible recombination of optical content with diverse infrared characteristics. Furthermore, we design a multi-level adaptive loss function that explicitly enforces complementary constraints on structural fidelity and semantic consistency during the translation process. Extensive experiments on two public datasets demonstrate the effectiveness of SSD-VI. On RGB-NIR, it achieves an FID of 46.53 and a KID of 0.0331, while increasing classification accuracy by 6.68 percentage points, from 83.37% to 90.05%. On VEDAI, SSD-VI improves mAP@50 by 0.13 for YOLOv8m and 0.14 for RT-DETR, confirming the value of the generated samples for infrared target recognition.

Journal of Imaging doi: 10.3390/jimaging12080352

Authors: Ian D. Li Choong-Yong Ung Cristina Correia

Accurate estimation of Alzheimer’s disease (AD) severity from structural magnetic resonance imaging (MRI) remains difficult, as disease-associated anatomical alterations are often subtle and publicly available datasets are typically too small to support robust deep learning model training. This feasibility study sought to determine how much Alzheimer’s disease spectrum-related information could be extracted from a small structural MRI cohort using a deliberately lightweight two-dimensional convolutional neural network (2D CNN), and whether transfer learning improves model performance. This study was intended as a methodological proof of concept rather than the development of a clinically deployable diagnostic tool. Structural scans and Clinical Dementia Rating (CDR) labels from the OASIS-1 dataset were filtered to 214 subjects: 124 cognitively normal (CN), 65 with mild cognitive impairment (MCI; CDR = 0.5), and 25 with AD-level impairment (CDR ≥ 1). A compact 2D CNN trained from scratch and a transfer learning model (frozen ImageNet MobileNetV2 features) were evaluated on four binary tasks (CN vs. AD, MCI vs. AD, CN vs. MCI, and CN vs. any impairment) under identical pre-processing and subject-level repeated 5-fold cross-validation (10 repeats), with the decision threshold tuned only on an inner split. Discrimination was summarized by ROC-AUC with 95% confidence intervals (CIs), permutation tests against chance, and per-task sensitivity and specificity. The from-scratch CNN recovered only a broad normal-versus-impaired signal (CN vs. any impairment AUC 0.59) and was at chance on adjacent-stage tasks (MCI vs. AD 0.41; CN vs. MCI 0.51). Transfer learning improved every task: CN vs. AD AUC 0.745 (95% CI 0.730–0.763), CN vs. any impairment 0.642, CN vs. MCI 0.601, and MCI vs. AD 0.599. On an independent OASIS-2 cohort, the transfer learning CN vs. AD model retained AUC 0.748. In this small-data regime, transfer learning recovers substantially more Alzheimer-spectrum signals than a from-scratch CNN, but performance remains modest because it is bounded by CDR-based, non-biomarker-confirmed labels, suggesting the model separates CDR-defined cognitive-status groups rather than detecting AD pathology.

Journal of Imaging doi: 10.3390/jimaging12080351

Authors: Arslan Munir

Image processing and computer vision continue to play transformative roles across science, engineering, healthcare, transportation, agriculture, manufacturing, environmental monitoring, and intelligent systems [...]

Journal of Imaging doi: 10.3390/jimaging12080350

Authors: Miaomiao Zhang Meng Lou Linwei Chen

Facial expression recognition in the wild is challenged by both noisy supervision and degraded visual evidence: subtle expression cues must be interpreted under blur, contrast changes, image noise, and annotator disagreement. Existing noisy-label FER methods mainly regulate samples, labels, or attention, while frequency information is rarely used to adapt the semantic representation itself. We propose Frequency-Guided Expert Modulation (FARM-FER), which treats local and global frequency descriptors as a control signal rather than an additional classifier input. A joint Haar-DWT and radial-FFT context guides soft routing among nonlinear experts and channel-wise affine recalibration of the semantic feature, while a learned gate combines the two corrections before a lightweight classifier predicts the expression from the refined representation. Across RAF-DB, FER+, and AffectNet under symmetric label noise, with additional evaluations under class-dependent label noise on RAF-DB and native crowd-label ambiguity on FER+, FARM-FER consistently improves matched baselines. At 30% symmetric noise, FARM-FER reaches 89.18% accuracy on RAF-DB, with a 1.6% performance gain over the matched Swin-Tiny baseline; the gains also hold in a controlled ResNet18 reimplementation and in class-sensitive AffectNet evaluation. Measured cost analyses show only modest parameter and FLOP overhead, supporting a lightweight yet effective design in terms of model size and arithmetic cost for noisy-label FER.

Journal of Imaging doi: 10.3390/jimaging12080349

Authors: Anatoly Sidorov Alexey Zaripov Ivan Tikshaev Vladislav Pozdyshev

In this paper, an approach to automating the monitoring of compliance with regulated technological operations using computer vision methods is proposed and investigated, with the loading of petroleum products considered as a case study. A distinctive feature of this work is the integration of a formalized description of the production process in BPMN notation with an object detection system, enabling not only object recognition but also interpretation of the sequence of technological actions performed by personnel. Based on the collected and annotated dataset containing more than 6000 images, a YOLOv11 neural network model was trained to monitor key stages of a technological operation. The experimental results show that the trained model provides high accuracy in detecting objects during the daytime (mAP50 is approximately 0.98), while maintaining the ability to work in real time. The results obtained confirm their applicability in industrial conditions. The work revealed the dependence of the quality of computer vision system functioning on the illumination conditions of the production area. It has been established that at night there is a significant decrease in recognition accuracy due to the presence of glare from lighting sources directed at the camera area. The results obtained make it possible to substantiate the need to take into account lighting factors when designing video monitoring systems for technological processes. To move from the level of object detection to monitoring compliance with regulations, an algorithm for interpreting detected objects has been developed, which ensures the fixation and analysis of the sequence of operations performed. Experimental tests conducted at the existing production site have confirmed the possibility of automated detection of violations of technological regulations and an increase in the level of industrial safety. The directions for further development of the proposed approach have also been identified, including the expansion of the training sample, taking into account a variety of production scenarios, and the development of methods to increase the stability of the system in difficult light conditions.

Journal of Imaging doi: 10.3390/jimaging12080348

Authors: Mohammed Aliy Mohammed Esla Timothy Anzaku Jef Jonkers Janarthanan Krishnamoorthy Wesley De Neve Sofie Van Hoecke

Conformal prediction has emerged as a principled framework for uncertainty quantification in computer vision, offering rigorous finite-sample coverage guarantees. However, its application in object detection has remained largely confined to localization, as standard inference codebases typically yield only top-1 class scores, precluding full class-label conformalization. In this work, we bridge this gap by adapting four architecturally diverse detectors—Faster R-CNN, RetinaNet, YOLO11, and RT-DETRv2—to facilitate the extraction of comprehensive per-class score vectors and the estimation of background confidence in the absence of native background modeling. Leveraging these adapted architectures, we implement inductive conformal prediction (ICP) using five distinct nonconformity functions: Top-K, Adaptive Prediction Sets (APS), Hinge, Margin, and Brier score. Our framework is rigorously benchmarked across a curated 20-class subset of MS-COCO and two specialized parasite egg datasets (AI4NTD P1.5v2 and Chula-ParasiteEgg-11). In addition, a Naive cumulative-threshold method is included as a baseline for comparison with APS, given their comparable mathematical formulations. Across target coverage levels of 90%, 95%, and 99%, the conformalized models consistently achieved nominal coverage with only minor finite-sample deviations. Hinge and APS exhibited an optimal balance between statistical coverage and prediction-set efficiency, whereas Margin and Brier scores tended toward larger sets under high data complexity and strict coverage requirements. With empty prediction sets maintained below 0.1%, our findings establish ICP as a robust and adaptable paradigm for trustworthy class-label uncertainty estimation, particularly within safety-critical workflows such as automated parasite diagnostics.

Journal of Imaging doi: 10.3390/jimaging12080347

Authors: Laila Rasmy Imane Sebari Mohamed Ettarid

This paper proposes a hybrid phase-correlation framework with a new multiscale detector, Fusion. Using a combination of FAST and Shi–Tomasi keypoints, followed by a probabilistic Hough transform and Canny edge detection, this detector improves repeatability. In addition, due to the limited ability of standard phase correlation to handle large geometric displacements, a complementary strategy is required to achieve sub-pixel matching precision. First, corners are extracted from both reference and sensed images using the Fusion detector. Corresponding points are then identified through coarse-to-fine phase correlation across a Gaussian pyramid. At each level, phase correlation yields an initial displacement, which is refined to sub-pixel accuracy using 1D parabolic fitting and propagated upward through the pyramid to obtain the final displacement. The proposed approach is evaluated using Pleiades and Sentinel-2 satellite images. Compared with the Scale-Invariant Feature Transform (SIFT)-based method and the detector-free Local Feature Transformer (LoFTR), the proposed framework achieves an RMSE below 0.2 and 0.4 pixels for Sentinel-2 and Pleiades imagery, respectively. Moreover, the results of the optimization analysis have revealed that shows that 2D paraboloid fitting combined achieves the lowest registration error of 0.010 pixels and the highest inlier ratio of 40.6%. The proposed approach achieves sub-pixel accuracy in the presence of noise and produces large numbers of correct matching points across different image resolutions.

Journal of Imaging doi: 10.3390/jimaging12080346

Authors: Taiwei Sun Lei Xue Tingting Li Shisuo Du Anning Cao Yang Shen Bei Lv Weixing Ji Ze Wang

Background: Ki-67 is a pivotal biomarker of tumor proliferative activity in esophageal cancer, yet its clinical application is hindered by reliance on invasive biopsy. Radiomics offers a non-invasive alternative, but conventional methods may be confounded by inter-individual baseline variations. This exploratory study aims to develop a radiomics-based biomarker for predicting Ki-67 expression. Methods: This single-center retrospective study included 59 patients with esophageal cancer. Delta-radiomics features were derived from preoperative CT images by calculating the difference between radiomic features from the tumor and paired normal esophageal tissue. Feature selection (mRMR, k = 3) was nested within leave-one-out cross-validation (LOOCV) to prevent data leakage. A Random Forest model was compared with Logistic Regression and Support Vector Machine across three feature types, five Ki-67 thresholds, and clinical variables. SHAP analysis was used for interpretability. Results: The Random Forest model achieved an AUC of 0.643 (95% CI: 0.483–0.792). Delta radiomics outperformed esotarget (AUC = 0.546) and eso (AUC = 0.514) models. The combined model (AUC = 0.619) did not outperform delta radiomics alone. SHAP analysis identified GrayLevelVariance and SmallAreaEmphasis as the most influential features. Conclusions: This exploratory study demonstrates that delta radiomics provides moderate discriminatory performance for predicting Ki-67 expression. External validation in independent multi-center cohorts is required before clinical application.

Journal of Imaging doi: 10.3390/jimaging12080345

Authors: Jiajia Lu Yue Shen Xu Wang Fuyang Ke

Aiming at the current SLAM (Simultaneous Localization and Mapping) algorithms in urban scenarios, which have problems such as elevation drift, odometry drift, and the appearance of false loop closures, a tightly coupled SLAM method with LiDAR and inertial guidance is proposed. In the front-end, a raster-based point cloud feature extraction method is introduced, enabling simultaneous segmentation and extraction of line, surface, and ground features. Utilizing the alignment results of line and surface features as the initial value for ground point alignment, interpolation weights are determined based on roll and pitch angle errors, effectively reducing global elevation errors through frame-by-frame constraints. The back-end employs an error state-based Kalman filter (ESKF) for GPS and IMU data fusion, enhancing the validity of true state estimation. A Scan Context loop closure detection method is designed, augmented by GPS detection as an auxiliary loop closure constraint to mitigate false loop closures. A global factor graph optimization model is also proposed. Experimental results demonstrate that, compared to existing open-source algorithms, the proposed method exhibits improved performance in structured urban scenes, reducing the average RMSE APE by 47.4% compared with LiDAR-only methods and by 22.9% compared with tightly coupled LiDAR-inertial methods. This work highlights the potential of multi-sensor fusion SLAM for achieving high-precision 3D localization and mapping in complex urban environments.

Journal of Imaging doi: 10.3390/jimaging12080344

Authors: Sreelekshmi Vijayasree Adithya Krishna Akarsh S. Nair Alfy Alex Shyamdev Krishnan Jayakrishnan Jyothisha J. Nair

Medical image analysis has undergone transformative progress with the application of deep learning models. However, existing architectures often struggle to effectively balance local feature extraction with global contextual understanding, which is crucial for complex diagnostic tasks such as Retinopathy of Prematurity (ROP) detection. In this study, we present a pretrained lightweight Conformer model tailored for medical image classification. The model integrates convolutional layers for capturing fine-grained spatial features with transformer blocks that capture long-range dependencies, creating a unified architecture capable of robust representation learning. We evaluate the model across multiple benchmark medical imaging datasets, including ROP, BloodMNIST, RetinalMNIST and other MedMNIST benchmark datasets. With 93.61% accuracy on the ROP dataset and 99.12% accuracy on BloodMNIST, experimental results show competitive classification performance while lowering model complexity to 12.4 million parameters and 3.2 GFLOPs. Experimental results demonstrate that the comparative studies versus CNN-based and transformer-based architectures, such as ResNet50, Swin-Tiny, ConvNeXt-Tiny, Vision Transformer, and MedViT. The findings show that in clinical settings with limited resources, the suggested lightweight Conformer offers a practical and computationally efficient alternative for medical image interpretation. Furthermore, the lightweight design ensures computational efficiency, making it suitable for deployment in resource-constrained healthcare environments. These findings validate the lightweight Conformer model’s potential for scalable, accurate, and real-time medical image classification.

Journal of Imaging doi: 10.3390/jimaging12080343

Authors: Yongkang Zhu Tianyue Yu Hongmei Li Xin Shu

Reliable medical image segmentation remains challenging because models must preserve fine boundary details while maintaining global semantic consistency. CNNs capture local structures effectively but have limited long-range modeling ability, whereas Transformer-based methods improve global context at high computational cost. Mamba-based state space models offer efficient long-range modeling, but may weaken high-frequency textures and boundary cues. To address these limitations, we propose DIG-MambaNet, a Dual-path Interactive Guided Mamba Network for medical image segmentation. The network introduces a dual-path complementary modeling block (DCM Block), where a cross-feature spatial interaction module (CSIM) adaptively integrates CNN-based local features and Mamba-based global features. A source image-guided module (SIGM) injects high-frequency information from the original image to compensate for downsampling-induced detail loss, while an inter-layer detail refinement fusion module (IDRFM) improves encoder–decoder feature alignment during reconstruction. Experiments on 2018DSB, ISIC2018, JSUAH-Cerebellum, and CVC-ClinicDB, covering nuclei segmentation in microscopy images, skin lesion segmentation in dermoscopic images, fetal cerebellum segmentation in ultrasound images, and polyp segmentation in colonoscopy images, demonstrate that DIG-MambaNet achieves consistent and competitive performance across diverse target structures and imaging conditions, with improved boundary delineation and favorable overlap-based accuracy compared with representative CNN-, Transformer-, and Mamba-based methods.

Journal of Imaging doi: 10.3390/jimaging12080342

Authors: Yang Zheng Wu Chen Zihao Xu Qingcang Yu

To address the difficulty in growth point localization caused by cotyledon overlapping and leaf occlusion during plug seedling grafting, a localization method based on grid structured light is proposed. An acquisition system consisting of a complementary metal-oxide-semiconductor (CMOS) camera and a grid structured light laser projector is established. A multi-depth plane calibration method is adopted to fit the light plane equation for each laser line. To address the grid line discontinuity problem, a coding method based on three-dimensional constraints of light planes is proposed. The cotyledon point cloud is reconstructed by combining the light plane equations with the camera model, and a circumscribed triangle is constructed to approximate the arc center of the fan-shaped point cloud for growth point localization. Experimental results show that the average interlayer error of light plane calibration is 0.059 mm. For 50 non-overlapping single seedlings, the average localization error is 1.68 mm with a success rate of 100%. For 150 overlapping seedlings, the success rate reaches 89%, outperforming the traditional ellipse fitting method (68%). The proposed method can provide reliable growth point localization information for grafting robots.

Journal of Imaging doi: 10.3390/jimaging12080341

Authors: Seyed Sina Ziaee Katie Ovens

Uncertainty quantification (UQ) in medical image segmentation is essential for ensuring the reliability and interpretability of deep learning models in clinical decision-making. While convolutional neural networks (CNNs) and transformer-based architectures have achieved remarkable segmentation performance, they often provide deterministic outputs without accounting for uncertainty, which can lead to overconfident predictions in ambiguous cases. This paper presents a comprehensive survey of UQ techniques in medical image segmentation, categorizing existing approaches into Bayesian methods, deep ensembles, deterministic methods, test-time data augmentation, and hybrid models, while treating foundation-model-based UQ as a separate cross-cutting category. We examine key methodologies, including Monte Carlo dropout, Bayesian neural networks, variational inference, and ensemble learning, discussing their advantages and limitations in addressing aleatoric and epistemic uncertainties. Additionally, we explore the clinical relevance of UQ by reviewing its applications in brain tumor segmentation, cardiac imaging, lung nodule detection, and other medical domains. The paper also highlights key evaluation metrics, such as calibration errors, uncertainty–error correlation, and visual interpretability, to assess the effectiveness of UQ methods. Finally, we discuss challenges and future research directions, emphasizing the need for scalable, interpretable, and clinically actionable uncertainty quantification strategies to improve trust in AI-assisted medical image analysis.

Journal of Imaging doi: 10.3390/jimaging12080340

Authors: Stojche Rechanoski Jasmina Veta Buralieva Saso Koceski Nikolay Hinov

The fusion of information from multiple imaging modalities plays a very important role in medical diagnosis. Wavelet-based transformations have been identified as powerful methods for this purpose due to their multiresolution nature, which enables the simultaneous preservation of both structural and fine-detail information across different frequency bands. In this paper we present experimental results obtained from the wavelet-based decomposition and fusion of medical images using Python and PyWavelets. Seven wavelets from four wavelet families—Daubechies (‘db1’, ‘db10’), biorthogonal (‘bior1.3’, ‘bior4.4’), coiflets (‘coif1’, ‘coif10’), and discrete Meyer (‘dmey’)—were systematically evaluated across three decomposition levels. An emphasis was put on the preservation of approximation and detail sub-images. Results outline that simpler wavelets used for the wavelet-based decomposition of grayscale medical images produce more details when compared with the colored medical images. From the tested fusion rules, and for the specific image pairs used in the analyses, we conclude that the average fusion rule gives the best information, without a lack of or excess of information regarding the visual quality of the fused image. Considering entropy as a quality metric and according to its higher values at all levels, the ‘bior4.4’ wavelet emerges as the best for wavelet-based image fusion. These findings could provide practical guidance for wavelet selection in multimodal medical image fusion pipelines.

Journal of Imaging doi: 10.3390/jimaging12080339

Authors: Zhengqi Liu Lijun Xu

State-space models segment images in linear time, but existing dermoscopic segmenters serialize the two-dimensional feature map along only one or two scan directions and operate purely in the spatial domain, which dilutes the orientation cues and fine boundary information that distinguish a pigmented lesion from surrounding skin. We address both limitations in a single linear-complexity network, Hydra-DermSeg-Net, that unifies three components not previously combined for this task: a four-directional bidirectional Hydra block that aggregates forward and backward selective scans over horizontal, vertical and two diagonal trajectories with learnable fusion weights; a differentiable discrete cosine transform (DCT) branch that decouples high- and low-frequency content so that boundary detail is processed separately from global semantics; and a learnable local contrast-enhancement front-end coupled with a clDice-supervised boundary attention gate. On a merged ISIC 2017/2018 corpus of 3994 images, the model attains a Dice coefficient of 0.9041 and an Intersection-over-Union (IoU) of 0.8386 against five baselines (U-Net, Att-UNet, VM-UNet, TransUNet and MALUNet) trained from scratch under a unified protocol, and it transfers to an unseen HAM10000 subset at 0.9347 Dice. It attains the highest Dice and IoU on each of the three data partitions examined, the highest Sensitivity, and the lowest 95-percentile Hausdorff distance under every random seed; the margin over the strongest convolutional baselines is about 0.002 in Dice and lies within the variation between training runs. An ablation at the full model capacity isolates the contribution of each component: removing the frequency branch costs 0.86 points of Dice and 0.99 points of Boundary IoU, and removing the boundary attention gate 0.66 and 2.85 points respectively. These results show that combining multi-directional state-space scanning with frequency-domain decoupling yields accurate and parameter-efficient segmentation without the quadratic cost of self-attention.

Journal of Imaging doi: 10.3390/jimaging12080338

Authors: Prashanth Gowda Attahalli Shivakumar Azhar Mahmood Shaheen Khatoon

Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, where early and accurate diagnosis is critical for effective treatment. Although recent advances in deep learning have enabled automated radiology report generation from breast ultrasound images, most existing approaches function as black-box systems, limiting clinical trust and interpretability. This study proposes a trustworthy and explainable framework for automated breast ultrasound report generation that combines Vision-Language Modelling (VLM) with multi-level Explainable Artificial Intelligence (XAI). The proposed architecture integrates a Swin Transformer for visual feature extraction, BioBERT/ClinicalBERT for clinical text representation, and a GPT-2-based decoder for report generation through a dual cross-attention fusion mechanism. The framework is evaluated on benchmark breast ultrasound datasets paired with expert-annotated radiology reports using standard natural language generation metrics, including BLEU, ROUGE-L, METEOR, and CIDEr. Experimental results demonstrate that the multimodal architecture significantly improves report quality, clinical consistency, and semantic accuracy compared with conventional image-only and single-modal baselines. To address transparency and trustworthiness, the framework provides dual-level explanations through Grad-CAM visual heatmaps and LIME/SHAP-based token attribution analysis, enabling clinicians to understand both image regions and textual features influencing generated reports. Qualitative assessment further indicates strong alignment between model explanations and radiologist-identified diagnostic findings.

Journal of Imaging doi: 10.3390/jimaging12080337

Authors: Fenna Brunken Benno Bullert Robert Brauweiler Paul A. Grützner Sven Y. Vetter Nils Beisemann

Different flat-panel detector technologies are available for intraoperative fluoroscopy. This study compared orthopedic and trauma surgeons’ preference and subjective image quality among mobile C-arm systems equipped with amorphous silicon (a-Si), complementary metal oxide semiconductor (CMOS) or indium gallium zinc oxide (IGZO) detectors. Fluoroscopic imaging was performed on four human specimens at four anatomic locations at pulse rates of 1/s and 10/s in low- and high-dose settings using three C-arm systems. Subjective image quality was rated by two observers on 5-point Likert scales. Pairwise forced-choice comparisons of images with identical acquisition parameters were analyzed using a Bradley–Terry model. Across all images, CMOS- and IGZO-based systems were preferred over the a-Si-based system in 91.8% and 90.4% of comparisons, respectively (ORs 9.53 and 11.20; both p < 0.001). No significant overall preference was observed between the IGZO- and CMOS-equipped systems. Subjective image quality ratings were significantly higher for CMOS- and IGZO-based systems compared with the a-Si-based system, particularly for overall image quality and perceived noise, while no consistent differences in image quality were found between CMOS- and IGZO-based systems. Overall, the CMOS- and IGZO-based systems evaluated in this study were preferred over the a-Si-based system and achieved superior subjective image quality ratings.

Journal of Imaging doi: 10.3390/jimaging12080336

Authors: Lanfeng Cui Yi Zong Fanqiang Kong Jianxin Li

This paper proposes a self-supervised polarization image dehazing method with an angle-of-polarization (AoP) frequency-domain prior for strong scattering dense-haze scenarios. The method formulates dehazing as the recovery of the clear object-radiance polarization field, rather than only restoring a haze-free intensity image. By analyzing real polarized hazy images, we observe that atmospheric AoP is dominated by low-frequency components, while object-radiance AoP contains richer local variations. Based on this observation, an AoP frequency-domain prior is incorporated into the polarization scattering model to guide the separation of object radiance and atmospheric polarization. A two-stage self-supervised training framework is then developed, where physical priors and the AoP prior provide stable component estimates, followed by joint optimization through scattering reconstruction consistency. In the object-radiance branch, a spatial-frequency dual-domain enhancement module is designed to capture both global haze degradation and local structural details. Experiments on a self-collected real short-wave infrared polarized hazy image dataset demonstrate that the proposed method achieves better target visibility, structural restoration, and quantitative performance than existing methods under dense-haze and strong scattering conditions.

Journal of Imaging doi: 10.3390/jimaging12080335

Authors: Alexandros Barmperis Olga Menegaki Anna Panagiotakopoulou Andreas Vezakis Ioannis Vezakis Ioannis Kakkos George K. Matsopoulos

Accurate segmentation of abdominal organs in Computed Tomography (CT) underpins radiotherapy planning, surgical planning, and disease monitoring. Existing benchmarks rank architectures by a single aggregate Dice score, without per-organ statistical testing or boundary-sensitive metrics, even though models are chosen organ by organ for clinical use. We benchmark ten architectures spanning convolutional, attention-based, transformer, and state–space (Mamba) families on the AMOS CT dataset under one identical nnU-Net-style pipeline; we report per-organ Dice, 95-percentile Hausdorff Distance (HD95), and Normalised Surface Dice, with pairwise significance tested on an independent external dataset (TotalSegmentator). A competitive cluster of convolutional and Mamba models leads; rankings are stable on large organs but reshuffle by 10–13% on the small, geometrically complex ones, and boundary fidelity separates the models into tiers that the Dice ranking hides. This ordering largely holds on the external set (Spearman ρ=0.84). Selecting a model on aggregate Dice alone is therefore unsafe for organ-specific clinical tasks: per-organ overlap and boundary metrics should be the primary acceptance criteria for selecting a model before clinical deployment.

Journal of Imaging doi: 10.3390/jimaging12070334

Authors: João Ferreira Nunes Pedro Miguel Moreira João Manuel R. S. Tavares

Human gait datasets play a central role in the development and evaluation of computer vision models. However, the current dataset landscape remains highly heterogeneous, with inconsistent reporting of acquisition conditions, user variability, and sensing configurations, which limits reproducibility and hinders principled cross-dataset comparability. In this work, we propose a covariate-centered, modality-agnostic taxonomy for gait datasets, explicitly structuring variability across scene-level, user-level, and sensor-level factors. The proposed framework enables consistent characterization of datasets through a standardized set of covariates (A–R), bridging differences across application domains and sensing modalities. Following a systematic review protocol aligned with PRISMA 2020, we analyze 47 publicly available image- and depth-based human gait datasets spanning healthcare, biometric, and attribute-recognition application domains. Using the proposed taxonomy, we derive a quantitative analysis of covariate coverage, revealing systematic biases in current dataset design.

Journal of Imaging doi: 10.3390/jimaging12070333

Authors: Ioannis Antonakos Emmanouil Anousis Tatiana Roko Antonia Alexiadou Maria Dimitropoulou Dimitris Filippiadis Stavros Spiliopoulos Konstantinos Palialexis Athanasios Giannakis Niki Parmenidou Efstathios Efstathopoulos

Angiography is a key imaging modality for the diagnosis and treatment of vascular diseases, and the growing sophistication of interventional procedures has heightened the need for radiation dose optimization. Diagnostic Reference Levels (DRLs) are widely used to monitor patient exposure and to support optimization in accordance with the ALARA principle. This study compared radiation dose metrics from a newly installed angiographic system at Attikon University Hospital with those obtained from the institution’s previous system and with values reported in the published literature. Radiation dose and procedural parameters were retrospectively collected for digital cerebral subtraction angiography (DSA), embolization, nephrostomy, vertebroplasty, transjugular intrahepatic portosystemic shunt (TIPS), chemoembolization, and injection procedures. Dose area product (DAP), fluoroscopy-related DAP, patient entrance dose indicators, and fluoroscopy time were analyzed. Median DAP values ranged from 5.72 to 349.60 Gy·cm2 depending on the procedure. Compared with data acquired approximately a decade earlier, DAP values increased by an average of 96.4%, whereas fluoroscopy times remained largely unchanged. Despite this increase, dose levels were generally lower than those reported in the international literature. Median DAP values ranged from 5.72 to 349.60 Gy·cm2 depending on the procedure. Compared with data acquired approximately a decade earlier, DAP values increased by an average of 96.4%, whereas fluoroscopy times remained largely unchanged. Although these differences may reflect the combined influence of technological developments, evolving procedural complexity, operator-related factors, and changes in clinical practice over time, these variables were not directly assessed in the present retrospective study. Nevertheless, the updated institutional Diagnostic Reference Levels provide a valuable benchmark for radiation dose optimization, quality assurance, and future multicenter studies aimed at supporting national DRL establishment.

Journal of Imaging doi: 10.3390/jimaging12070332

Authors: Yaqian Zhang Guanjun Wang Quan Zhang Bochao Zhou

In the restoration of mural images with rich colour information and complex texture structures, existing techniques typically extract the spatial-domain features in the Red–Green–Blue (RGB) colour space. However, the three RGB channels are physically decoupled without unified perceptual colour correlation constraints, which often leads to noticeable colour deviation in damaged regions with large colour variations. In addition, restoring both high-frequency texture details and low-frequency global structures in a mixed-frequency spatial domain can create conflicts between frequencies, making it difficult to generate realistic high-frequency details. To address these issues, we propose the laboratory frequency network (LABFNet), a restoration network guided by the laboratory (LAB) colour space and frequency-domain constraints. Our model has two key improvements: (1) it incorporates colour parameters from the LAB space to model colour loss in murals, and (2) it decomposes the image into low- and high-frequency components and enforces frequency consistency during restoration. In the Dunhuang 20–40% mask-ratio setting, the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) improved by 1.58% and 0.27%, respectively, while the Mean Absolute Error (MAE), Learned Perceptual Image Patch Similarity (LPIPS) and CIEDE2000 decreased by 5.87%, 6.5%, and 21.65%, respectively. Experimental results on benchmark datasets show that LABFNet reduces colour deviation and structural defects.

Journal of Imaging doi: 10.3390/jimaging12070331

Authors: Dirk Elias Schut Rachael Maree Wood Rob Schouten Robert van Liere Tristan van Leeuwen Kees Joost Batenburg

This study presents two workflows for leveraging longitudinal computed tomography (CT) datasets when developing deep learning-based detection systems for gradually developing postharvest disorders. Workflow 1 (Longitudinal Benchmarking) benchmarks neural networks by training and testing them on images from different stages of disorder progression. It examines the trade-off between detecting a disorder early or accurately and evaluates whether neural networks can generalize across time points. Workflow 2 (Longitudinal eXplainable Artificial Intelligence (XAI) Heatmaps) provides heatmaps that indicate how changes over time affect the outcomes of neural networks. It uses image registration to align an earlier-acquired image and then uses it as a baseline when calculating the heatmap. The workflows are demonstrated on a dataset of ‘Braeburn’ apples that were CT-scanned multiple times while developing internal browning during controlled-atmosphere (CA) storage and shelf life. The Longitudinal Benchmarking workflow was used to investigate whether images acquired immediately after CA storage can be used to predict the eventual browning after a shelf-life period, which is highly relevant in industrial practice. Moreover, the longitudinal XAI heatmaps avoided artifacts caused by out-of-distribution baselines or identical baseline regions, which occurred with conventional black or zero baselines.

Journal of Imaging doi: 10.3390/jimaging12070330

Authors: Ali A. A. Alghamdi

The growing demand for raw and processed scientific data has encouraged many researchers and research institutions to adopt an open-source data policy. At present, data accessibility is of paramount importance due to the growing demand for artificial intelligence (AI) and machine learning (ML) applications in various scientific fields, particularly medicine. Medium- to large-scale mammography datasets are widely used in breast cancer research to develop and evaluate computer-aided detection methods. However, there are only a few studies on using mammogram datasets for the prediction of the breast mean glandular dose (MGD) with AI or ML models. The aim of this study was to investigate the feasibility of using ML and deep ML for MGD prediction based on DICOM images and retrieved dosimetric data from DICOM mammogram images. A total of 26,988 mammography images in DICOM format were obtained from the Federated Research Data Repository (FRDR). Eleven regression algorithms and three neural network-based models were evaluated using five-fold cross-validation. In addition, a deep ML fusion model based on Vision Transformer (ViT) and tabular data was developed for the prediction of the MGD normalized conversion factor CF(DgN). A mean breast thickness of 61.37 mm and a mean MGD of 1.53 mGy (0.55–6.33 mGy) were calculated using this dataset. Regarding tabular data, the artificial neural network (ANN) sequential models outperformed other linear and tree-based models. The ViT deep ML fusion model was tested with three configuration versions differing on the number of features included. A comparison of the three versions revealed that the version with six features achieved the best overall predictor performance. This study demonstrates that ML and deep ML can effectively predict the MGD using dosimetric tabular data and mammography DICOM images. The use of ML with tabular data extracted from DICOM images can be further strengthened by incorporating larger and more diverse datasets.

Journal of Imaging doi: 10.3390/jimaging12070329

Authors: Shengqun Zhang Annie Anak Joseph Kho Lee Chin

Pulmonary nodules are circular or irregular lesions visible on chest computed tomography (CT), and their early detection is critical for lung cancer screening. Deep learning detection algorithms have been widely adopted for pulmonary nodule diagnosis; existing lightweight models suffer from redundant network parameters and low detection accuracy for tiny lesions. To address these limitations, this study proposes an improved detection model based on YOLOv8n. First, Omni-Dimensional Dynamic Convolution (ODConv) replaces static convolution in the backbone to enhance multi-morphology nodule feature extraction. Second, the Convolutional Block Attention Module (CBAM) is embedded at multiple positions of the neck network to suppress background interference from blood vessels and normal lung parenchyma. Third, Complete Intersection over Union (CIoU) loss is substituted by Wise Intersection over Union (W-IoU) to optimize bounding box regression for hard samples with blurred boundaries. Experiments on the LUNA16 dataset show that compared with the original YOLOv8n, the proposed model improves Precision by 6.3%, Recall by 8.6%, mAP50 by 3.4%, and mAP50-95% by 2.7% while maintaining high inference speed. Additional generalization verification on the LIDC-IDRI multi-center dataset further proves the robustness of the proposed lightweight architecture, which achieves balanced accuracy and real-time performance compared with mainstream detection models.

Journal of Imaging doi: 10.3390/jimaging12070328

Authors: Raja Vavekanand Ganesh Kumar Muhammad Moazzam Jawaid Shafiya Qadeer Memon Teerath Kumar

Autism Spectrum Disorder (ASD) assessment remains challenging because behavioural instruments are partly observer-dependent and neuroimaging data are heterogeneous. This paper presents FAA (Fuse After Aligned), which is a multimodal classification framework that combines transfer learning for structural MRI (sMRI) representation learning with a contrastive objective for the pre-fusion alignment of sMRI and resting-state functional MRI-derived functional connectivity (FC) features. Evaluation was restricted to the single-site ABIDE-I New York University subset comprising 75 participants with ASD and 98 typically developing controls. Under the reported five-fold internal cross-validation protocol, FAA achieved a mean accuracy of 92.6% compared with 87.4% for naive fusion and 90.9% for the sMRI-only baseline. Ablation analyses indicate that adding the contrastive objective is associated with improved classification performance and that ResNet-18 outperforms the evaluated ViT-16 configurations in this small-sample setting. These findings support the methodological value of pre-fusion feature alignment within the evaluated cohort. The framework offers a robust, computationally efficient, and clinically viable approach for objective ASD diagnosis with strong potential for generalisation to multi-site neuroimaging applications.

Journal of Imaging doi: 10.3390/jimaging12070327

Authors: Ignacio Díaz-Lorenzo Rio Jorge Aguilar Torres Paloma Caballero Sanchez-Robles Raquel Caminero Garcia Alfonso Canabal Berlanga Alfonsa Friera Reyes Alberto Alonso-Burgos

(1) Objectives: In acute pulmonary embolism (PE), detailed thrombus burden scores are often complex and time-consuming, limiting their integration into urgent radiology reports. We evaluated a simplified modified Ghanima score (GmScore and GmS) designed to provide a structured estimate of thrombus burden and assessed its clinical correlation and reproducibility. (2) Methods: In this retrospective single-center study, 132 consecutive patients with confirmed acute PE were classified according to the modified GmScore: GmS1 (segmental), GmS2 (lobar), and GmS3 (main pulmonary arteries), considering luminal obstruction ≥ 50%. European Society of Cardiology (ESC) risk category, simplified Pulmonary Embolism Severity Index (sPESI), CT right-to-left ventricular (RV/LV) ratio, echocardiographic right ventricular dysfunction, and 30-day mortality were recorded. Inter- and intraobserver agreement were assessed using weighted kappa. (3) Results: In 132 patients (mean age 64.8 ± 16.5 years; 77 men), a significant clinical gradient was observed across GmScore categories. ESC intermediate–high/high risk occurred in 0% of GmS1 and 95.6% of GmS2–3 patients (p < 0.001). The median RV/LV ratio increased progressively (0.76, 1.58, and 1.79 for GmS1–3; p < 0.001), with a strong correlation between the GmScore and RV/LV (Spearman ρ = 0.75). GmS2 and GmS3 showed no significant difference in ventricular repercussion (p = 0.938), whereas GmS1 differed markedly. Using GmS ≥ 2 to identify ESC intermediate–high/high risk yielded 100% sensitivity and negative predictive value. Interobserver agreement was excellent (κ = 0.92). Thirty-day mortality was 0% in GmS1, 2.0% in GmS2, and 14.6% in GmS3 (p = 0.005). (4) Conclusions: The modified GmScore is a simple, reproducible CT-based descriptor that aligns closely with right ventricular repercussion and ESC risk stratification.

Journal of Imaging doi: 10.3390/jimaging12070326

Authors: Abror Shavkatovich Buriboev Akhram Nishanov Shuxrat Isroilov Inomjon Narzullaev Umidjon Djumayozov Shavkat Buriboyev Temur Azamov Parda Yuldashov Davron Shodmonov Djamshid Sultanov Abbos Abduvaytov

Accurate classification of pollen grains in microscopic images remains challenging because of noise, structural variability, background complexity, weak texture, and intra-class similarity. To address these issues, this study proposes a hybrid framework that integrates contour-signal modeling, spectral–wavelet analysis, and deep learning for robust microscopic pollen image recognition. In the proposed approach, microscopic pollen images are first converted into contour-based point-signal representations, allowing object boundaries to be analyzed as structured one-dimensional signals. To improve signal quality under real imaging conditions, the framework incorporates Gaussian, median, and contour-aware filtering together with defect-point detection and correction. The processed contour signals are then analyzed using Fourier transform, continuous wavelet transform, and discrete wavelet transform to extract complementary global and local descriptors. These enriched representations are provided to a convolutional neural network for final classification. Experiments conducted on a seven-class microscopic pollen-image dataset demonstrate that the proposed method outperforms conventional computer-vision and baseline deep-learning approaches. The best-performing hybrid configuration achieved an error rate of 6.4%, while the overall classification accuracy reached 0.977 with an F1-score of 0.966, compared with 0.837 for a traditional computer-vision pipeline. These results confirm that combining contour-based signal processing with hierarchical deep feature learning provides an effective and noise-robust strategy for microscopic pollen image recognition. However, the present validation is limited to pollen images, and further experiments on broader microscopic object datasets are required to assess generalization to other micro-object categories such as nanoparticles, fibers, rods, and synthetic microstructures.

Journal of Imaging doi: 10.3390/jimaging12070325

Authors: Serhat Dikyar Behcet Ugur Toreyin

The proliferation of high-definition video data necessitates highly efficient processing pipelines for real-time edge analytics. However, traditional object detection architectures rely exclusively on pixel-domain inputs, which renders the computationally prohibitive decoding phase a latency bottleneck. In this paper, we propose a novel dual-phase framework designed to achieve fast and efficient object detection directly within the partially decoded compressed-domain data. First, we introduce a partial decoding paradigm featuring the Low-Frequency Spectral Prioritization method on the encoder side. By systematically discarding high-frequency residual coefficients and retaining only a sparse subset of fundamental spatial frequencies, this method dramatically reduces transmission payloads and accelerates the standard decoding process. Second, to recover the structural fidelity lost due to the intentional omission of residual data, we employ a multi-granularity cross-domain knowledge distillation architecture. This strategy aligns global contextual features, foreground boundary attention maps, and final response logits, transferring rich representational capacities from a high-performing pixel-domain teacher network to a lightweight compressed-domain student network. Comprehensive experiments utilizing RetinaNet, FCOS, and GFL object detection networks on the COCO-mini dataset demonstrate the superiority of the proposed framework. By retaining fundamental residual coefficients within the HEVC pipeline, the proposed method reduces average decoding latency while improving the mAP score by +0.96% over the conventional fully decoded pixel-domain baseline on the COCO-mini dataset.

Journal of Imaging doi: 10.3390/jimaging12070324

Authors: Duc Tri Phan Hong Duc Nguyen

Egocentric multi-view image analysis refers to the processing of utilizing synchronized video streams captured from multiple wearable cameras worn on the head or body, providing complementary first-person perspectives of dynamic, real-world interactions. Unlike single-view egocentric vision, which may suffer from severe occlusions, motion blur, and limited field-of-view or traditional fixed-camera multi-view setups (assuming static geometry and controlled environments), egocentric multi-view systems leverage body-worn rigs to enable a more robust and flexible 3D understanding in open-world, mobile scenarios. In this work, we present a systematic survey of advancements in cross-view feature fusion, geometric consistency enforcement, open-world detection, human–object interaction (HOI) modeling, action segmentation, 3D reconstruction, and novel-view synthesis specifically tailored to wearable multi-camera platforms. Key datasets released between 2024 and 2026—including HOT3D (833 min of synchronized multi-view hand/object interactions from Project Aria and Quest 3), MultiEgo (first multi-egocentric dataset for 4D social scene reconstruction), and Ego-1K (large-scale 12-camera rig for dynamic 3D video synthesis) are thoroughly examined alongside an analysis of integrations with large language models (LLMs) and vision–language models that drive performance gains, typically in the 15–30% range over single-view baselines in hand tracking, HOI recognition, and reconstruction fidelity, although we show through a consolidated meta-analysis that this gain is task-dependent: larger for geometry-bottlenecked tasks such as in-hand object lifting, and smaller, method-dependent, or occasionally negative for semantic-recognition tasks such as keystep recognition under naive view fusion. These methods cover work in multi-view stereo, cross-view learning, and novel-view synthesis while addressing several real-time wearable constraints. Practical applications such as immersive Augmented Reality/Virtual Reality (AR/VR), assistive robotics, and healthcare monitoring are also discussed together with the challenges in motion calibration, benchmark diversity, and edge deployment ability. Thus, in this review, we attempt to fill a critical gap by focusing exclusively on wearable multi-view systems in an open-world setting, synthesizing the latest literature to chart future directions toward more embodied and continual learning agents.

Journal of Imaging doi: 10.3390/jimaging12070323

Authors: Hediyeh Toufani Richard M. Dansereau Philippe Phan Jefferson R. Wilson Eve C. Tsai

Accurate spinal cord segmentation is important for quantitative analysis of spinal cord magnetic resonance imaging, including measurement of cross-sectional area and diffusion-based microstructural characterization. In pathological conditions like cervical myelopathy, the shape deformation induced by cord compression is extreme, rendering automated segmentation particularly challenging. While deep learning-based methods yield good results in healthy or mildly pathological cases, their reliability suffers when anatomical assumptions fail under compression. In this work, we introduce a pathology-aware, boundary-focused preprocessing framework that directly aims to mitigate failure modes imposed by cord compression. Instead of generic preprocessing, each component aims to enhance intensity homogeneity, suppress noise and improve boundary visibility. At the core of this approach is a multi-representation input derived from a single T2*-weighted scan, whereby complementary intensity-, contrast- and edge-enhanced representations are fed to the U-Net model. The proposed framework is evaluated on spinal cord MRI data from three clinical centers (194 cervical myelopathy cases). The results demonstrate that the proposed preprocessing framework improves segmentation accuracy, robustness, and stability, particularly in anatomically challenging regions affected by compression. These findings highlight the importance of pathology-aware preprocessing for reliable spinal cord segmentation in cervical myelopathy.

Journal of Imaging doi: 10.3390/jimaging12070322

Authors: Zurifa Ajanović Saleha Redžepi Uzeir Ajanović Naida Spahović Amina Zorlak-Čavčić Emina Dervišević Admir Terzić Mirza Pojskić

Sex estimation is a fundamental component of biological profiling in forensic anthropology, particularly when skeletal remains are incomplete or fragmented. This study aimed to evaluate sex estimation of the cranial base using geometric morphometrics and to assess the predictive value of cranial base morphology for sex estimation. The study included 211 adult skulls (139 male, 72 female) from the Bosnian population. Each skull was digitized to generate 3D models, and 27 anatomical landmarks were recorded. Landmark coordinates were standardized using Generalized Procrustes Analysis, Principal Component Analysis, Discriminant Function Analysis with permutation testing, and regression of shape on centroid size. Statistically significant sex estimation was observed at both the form (shape and size) and shape levels. Classification accuracy based on cranial base form reached 92.81% for males and 86.11% for females. Shape-based classification, after removal of size effects, also showed high accuracy (90.65% for males and 81.94% for females). Regression analysis indicated that size contributed significantly but modestly to shape variation. The cranial base exhibits stable sexually dimorphic patterns and may represent a reliable anatomical region for sex estimation. These findings contribute to population-specific standards for the Bosnia and Herzegovina population and support the forensic applicability of 3D geometric morphometric approaches.

Journal of Imaging doi: 10.3390/jimaging12070321

Authors: Seyed Mohammad Alizadeh Henning Müller Mohammad Sadegh Helfroush

Retrieving histopathological images can assist in the recognition and treatment planning of several diseases. Nevertheless, high-dimensional features can make this process complex and inefficient. These challenges can be addressed by encoding the feature domain into binary codes of different lengths utilizing deep hashing approaches. Still, the vanishing gradient challenge remains a concern in these approaches. According to several studies, quadruplet deep hashing models have exhibited promising performance in retrieving images from multi-category datasets. Furthermore, adding an attention module to a convolutional neural network architecture can increase the efficiency of feature extraction. Thus, we introduce an adaptive quadruplet deep hashing model to retrieve histopathological images. Four designed deep hashing models with matching structures and parameters are utilized to produce hash codes. The resulting codes are trained according to a novel adaptive quadruplet loss function. The adaptive structure is capable of improving retrieval performance. The presented approach also suggests a novel hash layer for the vanishing gradient issue. In addition, a simple yet effective attention module is implemented to enhance feature extraction performance. Our model is evaluated on three publicly available histopathology datasets: Kather, Kimia Path960, and Kimia Path24C. The results indicate that the suggested approach achieves the highest mean average precision (MAP) of approximately 0.9940, 0.9983, and 0.9968 for the respective datasets. Based on experiments performed on the datasets, our model surpasses current hashing techniques.

Journal of Imaging doi: 10.3390/jimaging12070320

Authors: Qidong Han Shuo Feng Yang Gao Mengyao Li Teng Meng Ke Li Yuhao Yang

Lane detection is a critical component of autonomous driving systems, requiring both high accuracy and real-time performance under complex driving scenarios. Unlike current methods that rely on predefined lane counts, instance segmentation methods can handle an arbitrary number of lanes, making them more adaptable in real-world applications. However, this flexibility typically relies on dense pixel-level predictions, which necessitate large-scale networks and result in prohibitively high computational costs, hindering deployment on embedded platforms. To address these challenges, we present Fast-CenLaneNet, a lightweight architecture that improves inference efficiency while maintaining detection accuracy. Specifically, we design a lightweight backbone to reduce model parameters and computational cost, propose a learnable spatial similarity attention module to capture spatial dependencies within lane regions and enhance feature discriminability, and construct multi-branch output heads with Ghost convolutions to refine lane-related features with low computational overhead. Experiments on the TuSimple and CULane benchmarks demonstrate that Fast-CenLaneNet achieves a favorable accuracy–efficiency trade-off. On TuSimple, Fast-CenLaneNet obtains 96.40 ± 0.06% accuracy and 162.7 ± 6.8 FPS with 4.7 M parameters and 9.9 GFLOPs. Compared with CenLaneNet, it reduces the number of parameters by 89.1% and improves forward inference speed by 107.5%, with an accuracy decrease of only 0.08 percentage points.

Journal of Imaging doi: 10.3390/jimaging12070319

Authors: Emil Matthisson Hanns-Christian Breit Markus Obmann Jakob Wasserthal Martin Segeroth Daniel Boll

Objectives: To evaluate the impact of contrast phase, age, and sex on CT-derived body composition metrics—specifically attenuation and volume of subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), liver, and skeletal muscle. The potential of the proportion of muscle voxels below 0 Hounsfield units (HU) as a surrogate for fatty infiltration was also explored. Materials and Methods: A retrospective analysis of 866 multiphasic abdominal CT scans (non-enhanced [NE], arterial [ART], portal venous [PV]) from 2012 to 2022 was performed. Segmentation of SAT, VAT, liver, and skeletal muscle was conducted using the AI-based TotalSegmentator. Wilcoxon signed-rank tests and Bland–Altman analysis (mean bias and 95% limits of agreement) were applied to assess contrast-related effects; Spearman’s correlation coefficient was used to assess demographic associations. Results: Significant variation in attenuation and volume of SAT, VAT, and muscle was observed across contrast phases (p < 0.001). SAT attenuation was higher in NE and PV than in ART, while VAT attenuation was highest in PV. SAT volume increased and VAT volume decreased in contrast-enhanced phases. Attenuation and volume showed strong inter-phase correlation (ρ > 0.9). VAT attenuation was significantly higher in females, whereas VAT volume was significantly greater in males. VAT volume negatively correlated with liver attenuation (ρ = −0.33). Muscle voxels <0 HU were significantly reduced in contrast-enhanced scans. Conclusions: Contrast phase, age, and sex significantly influence CT-based body composition parameters. These confounding factors should be considered when using quantitative imaging biomarkers in clinical and research settings.

Journal of Imaging doi: 10.3390/jimaging12070318

Authors: Adelina Filofteia Ghilencea Daniel Octavian Costache Constantin Căruntu Maria Moga Raluca Simona Costache

Introduction. Psoriasis is a chronic inflammatory disease histologically characterized by epidermal hyperproliferation, altered keratinocyte differentiation and dermal vascular remodeling. Although the diagnosis is mainly clinical, non-invasive imaging methods, such as high-frequency skin ultrasound, allow an objective assessment of skin changes and disease activity. Material and Methods. We conducted a pilot, observational, cross-sectional and comparative study, conducted within the Dermatovenerology Department of the Central Military Emergency Hospital “Dr. Carol Davila”, Bucharest, which included 40 patients diagnosed with psoriasis vulgaris, of whom 22 received conventional systemic treatment (methotrexate 15 mg/week), and 18 received biological therapy. For each patient, a representative, clinically active and recently appeared psoriatic plaque was evaluated with ultrasound, and the thickness of the epidermis, the thickness of the dermis, the thickness of the hypoechoic subepidermal band (SLEB) and the Doppler signal were analyzed. Statistical analysis was performed using SPSS v26. Results. Patients under biologic therapy had significantly lower ultrasound parameters compared to those under conventional systemic therapy, especially regarding epidermis thickness, hypoechoic subepidermal band thickness and Doppler signal. The PASI score was significantly higher in the conventionally treated group. Also, significant positive correlations were found between the PASI score and the hypoechoic subepidermal band thickness and the Doppler signal, Conclusions. Ultrasound parameters represent useful objective markers in the evaluation of psoriasis, reflecting disease activity. Patients under biologic therapy presented, at the time of evaluation, imaging parameters suggestive of reduced skin inflammation compared to those treated conventionally.

Journal of Imaging doi: 10.3390/jimaging12070317

Authors: Dănuţ-Ovidiu Pop Adrian-Silviu Roman

Pedestrian detection is a fundamental component of Advanced Driver Assistance Systems (ADAS) and plays a key role in collision avoidance and the safety of vulnerable road users. This paper presents a structured review of pedestrian detection methodologies developed between 2000 and 2025, spanning classical vision techniques and modern deep learning architectures. We organize the review into two phases. First, we examine classical methods, including Histogram of Oriented Gradients (HOG)+Support Vector Machine (SVM), Viola–Jones, Deformable Part Models, and Integral Channel Features, which established the conceptual foundations of the field. Then, we analyze state-of-the-art deep learning architectures, categorized by detector stage (one-stage vs. two-stage), localization strategy (anchor-based vs. anchor-free), feature extraction paradigm (Convolutional Neural Network (CNN)-based vs. transformer-based), output representation (bounding box vs. instance segmentation), and computational profile (lightweight vs. heavyweight). Several design principles introduced by classical methods remain visible in modern architectures, indicating that they were not fully superseded. The review also examines publicly available benchmark datasets and compares the strengths and limitations of camera-, Light Detection And Ranging (LiDAR)-, radar-, and multi-sensor-fusion-based systems for ADAS deployment. We close by identifying six open problems for the field: adversarial robustness, real-time inference under embedded constraints, detection under adverse weather, dataset bias and demographic fairness, the deployment of Bird’s-Eye View (BEV) and unified perception on automotive hardware, and explainability for safety-critical use.

Journal of Imaging doi: 10.3390/jimaging12070316

Authors: Catherine Aurelia Christie Alexander Vasileios Magoulianitis Jiaxin Yang C.-C. Jay Kuo

Nuclei segmentation is a key task in digital histopathology, highlighting important aspects of nuclear morphology and topology in many cancer-related evaluations and studies. Variability in nuclear appearance both within and across different organs, stain heterogeneity, and inconsistencies in acquisition procedures contribute to the complexity of the task. The existing nuclei segmentation methods apply deep learning to address these challenges, using models with millions of parameters, thereby significantly increasing computational complexity. They also face limitations in generalizing to unseen organs and slide preparations. In this paper, we propose a transparent and lightweight Green U-Shaped Learning model for nuclei segmentation (NS-GUSL). NS-GUSL features a multi-scale architecture for coarse-to-fine refinement of probability maps, which are subsequently binarized using a novel low-confidence sample binarization (LCSB) technique. The model features a modular, feed-forward feature learning scheme with unsupervised representation learning and supervised feature selection and generation. A final morphological post-processing step refines the segmentation maps to improve instance separation while preserving nuclei convexity. The model was trained and tested on the MoNuSeg dataset and compared against other deep learning baselines for segmentation performance. In addition, external validation experiments were conducted to evaluate the proposed model’s generalizability to unseen organs and staining procedures. NS-GUSL exhibits the best panoptic segmentation performance and competitive detection quality across all datasets. Moreover, our model is shown to be compact, low in computational complexity, and to have a minimal carbon footprint, compared to other deep learning models, making it a suitable choice for deployment on edge devices.

Journal of Imaging doi: 10.3390/jimaging12070315

Authors: Fabrizio Urraro Nicoletta Giordano Vittorio Patanè Marco Piscopo Giovanni Ciani Giovanni Balestrucci Maria Chiara Brunese Anna Russo Mario Sansone Alfonso Reginelli

Introduction: Small hyperechoic renal masses are frequently detected incidentally on conventional ultrasound and are often presumed to represent benign lesions, particularly angiomyolipomas. However, malignant renal tumors, including renal cell carcinoma, may also appear hyperechoic when small, creating a diagnostic challenge at first-line imaging. This study aimed to evaluate the feasibility and exploratory diagnostic performance of B-mode ultrasound radiomics for differentiating benign and malignant small hyperechoic renal masses. Methods: This retrospective single-center study included adult patients with incidentally detected small hyperechoic renal masses measuring ≤3 cm and examined between July 2022 and April 2025. All lesions underwent standardized B-mode ultrasound assessment and multidisciplinary review. Final diagnosis was established by histopathology when available or by longitudinal ultrasound follow-up stability for lesions considered benign. Lesions were manually segmented on representative B-mode DICOM images, and original radiomic features were extracted using PyRadiomics version 3.0 according to standardized definitions compatible with the Image Biomarker Standardisation Initiative framework. A total of 114 original radiomic features were extracted from each lesion. The primary comparison was benign versus malignant lesions. Diagnostic performance was assessed using feature-level receiver operating characteristic analysis. Results: Forty-two lesions were included in the final radiomic cohort, including 26 malignant renal cell carcinomas and 16 benign angiomyolipomas. Malignant lesions included papillary renal cell carcinoma, chromophobe renal cell carcinoma, and clear-cell renal cell carcinoma. All malignant lesions were histologically confirmed. Among benign lesions, 14 angiomyolipomas were classified based on longitudinal ultrasound stability, whereas 2 were confirmed by ultrasound-guided percutaneous biopsy after mild dimensional increase during imaging surveillance. Among the extracted radiomic features, firstorder_Variance and firstorder_MeanAbsoluteDeviation showed the highest exploratory discriminatory performance, each achieving an area under the receiver operating characteristic curve of 0.837. Both features are first-order measures of gray-level dispersion within the segmented lesion. Higher values were observed in malignant lesions, suggesting greater intralesional grayscale heterogeneity compared with benign angiomyolipomas. Conclusions: B-mode ultrasound radiomics is feasible for the quantitative assessment of small hyperechoic renal masses and may provide complementary information for differentiating benign angiomyolipomas from malignant renal cell carcinomas. firstorder_Variance emerged as a representative candidate imaging biomarker of grayscale dispersion, with firstorder_MeanAbsoluteDeviation showing concordant performance as a related dispersion measure. These findings should be considered preliminary and hypothesis-generating and require validation in larger multicenter cohorts before clinical implementation.

Journal of Imaging doi: 10.3390/jimaging12070314

Authors: Kuanysh Nazarov Veronica Smirnova Murat Kenessarin Yekaterina Dubyagina Yeldos Kariyev Sergey Kichanov Ayazhan Zhomartova Bagdaulet Mukhametuly Elmira Myrzabekova

We studied the structural features in internal pores and mineral inclusions of several vessels from the Shubarat-1 archeological site in the Republic of Kazakhstan, dating to the late first millennium BC, using neutron tomography. Differences in the neutron attenuation coefficients of the constituent elements of pottery objects, as well as the high penetration capability of neutron tomography, make it possible to conduct non-destructive studies of rare ceramic vessels. By analyzing the three-dimensional tomography data, we can reconstruct the size and morphological parameters of internal pores and minerals. Based on these structural findings, we can clarify past pottery production processes.

Journal of Imaging doi: 10.3390/jimaging12070313

Authors: Mario J. Noh-Burgos Juan B. Chalé-Dzul Leticia Olivera-Castillo César Puerto-Castillo Nina Méndez-Domínguez Rosa E. Moo-Puc

Magnetic resonance cholangiopancreatography (MRCP) is widely used for biliopancreatic imaging; however, hyperintense gastrointestinal fluids in heavily T2-weighted sequences may interfere with visualization of the biliary and pancreatic ducts. Natural manganese-containing beverages have been investigated in MRCP-related imaging contexts, and yerba mate (Ilex paraguariensis A. St.-Hil.) has been studied to this end. However, its concentration- and sequence-dependent signal behavior under MRCP-like phantom conditions remains insufficiently characterized. This preclinical phantom study evaluated the concentration- and sequence-dependent MRI signal intensity behavior of an aqueous extract of Ilex paraguariensis. The extract was characterized by means of elemental analysis, total manganese and iron quantification, total phenolic content, antioxidant capacity, and LC-ESI-MS analysis. MRI phantom experiments were run at different extract concentrations using T1-weighted, T2-weighted, and single-shot turbo spin echo (SSHTSE) sequences. The dried extract contained 1.22 ± 0.04 mg/g total manganese and 0.40 ± 0.01 mg/g total iron. Calculated total Mn concentrations in phantom dilutions ranged from 0.06 to 0.97 mg/dL. The extract showed concentration- and sequence-dependent signal behavior, with T1-weighted signal enhancement and progressive signal suppression in T2-weighted and SSHTSE sequences. No T1/T2 mapping or r1/r2 relaxivity measurements were performed. LC-ESI-MS identified MS1-based putatively assigned phenolic features without MS/MS confirmation of extract peaks. Ilex paraguariensis aqueous extract showed preliminary concentration- and sequence-dependent MRI signal intensity changes under phantom conditions, including signal suppression in MRCP-like heavily T2-weighted sequences. These findings do not establish clinical applicability, safety, tolerability, comparative efficacy, or improved duct visualization. Further studies are needed, incorporating relaxometric measurements, comparator agents, formulation assessment, in vivo evaluation, and clinical validation.

Journal of Imaging doi: 10.3390/jimaging12070312

Authors: Yange Sun Mengdi Wang Chenglong Xu Huaping Guo Li Zhang Hongzhou Yue Yan Feng

Accurate identification of surface defects on steel strips is critical for manufacturing quality assurance and operational reliability. Although deep learning has greatly advanced defect detection, precise recognition remains challenging due to significant background texture interference, loss of spatial details, and semantic imbalance across multiscale features. To address these challenges, we propose A2S2C-Det, a novel detector that integrates dual-path adaptive aggregation with spatial–semantic compensation to enhance feature representation for defect detection. First, we design a plug-and-play semantic refinement bottleneck (SRB) that augments backbone features through multiscale perception and a feature-screening bottleneck, enabling the model to suppress background interference while capturing subtle defect shapes. We further introduce a dual-path adaptive aggregation (DPAA) module that fuses complementary information from cross-level semantic consistency and fine-grained structural cues via two coordinated pathways, alleviating semantic imbalance across scales. Finally, we develop a spatial-semantic gated compensation (SSGC) module that adaptively supplies semantic information to low-level features while delivering spatial details to high-level features, recovering lost spatial details in high-level features. Extensive experiments on three benchmark datasets demonstrate that our A2S2C-Det achieves mAP50 of 82.0%, 73.0%, and 91.2%, and mAP of 47.1%, 36.5%, and 60.6%, respectively, comparing favorably against current state-of-the-art methods.

Journal of Imaging doi: 10.3390/jimaging12070311

Authors: Mihai Fieraru

Fitness motions present some of the most challenging cases for monocular 3D human reconstruction: extreme articulations, heavy self-occlusion, and frequent self-contact. The Fit3D dataset captures these motions at large scale via a 12-camera VICON motion capture system synchronized with 4 RGB cameras, but in its original release provides only 3D skeletons. This paper contributes three studies built on top of Fit3D. First, we design and validate a methodology for constructing dense GHUM and SMPL-X pseudo-ground-truth shape and pose annotations on top of the raw MoCap: an optimization-based fitting pipeline that combines markers, multi-view 2D keypoints, separate body and hand normalizing-flow priors, and a self-collision loss, which we show improves on the marker-only MoSh++ baseline on the hands and extremities. Second, we define a standardized evaluation protocol—metric set, frame sampling, and coordinate conventions—for monocular 3D human reconstruction on Fit3D, served through the IMAR-hosted Fit3D evaluation resource, and use it to conduct a comparative benchmark study of 19 representative methods spanning optimization-based, single-frame, and video-based families; the two trained on Fit3D (NLF and SMPLest-X) lead the position and orientation metrics, respectively. Third, a controlled fine-tuning experiment shows that adding Fit3D to the training mixture of a strong baseline (HMR2.0) sharply lowers error on the hardest fitness poses without degrading out-of-domain generalization. The Fit3D dataset and the GHUM/SMPL-X annotations are available, under a non-commercial research license, through the IMAR Fit3D resource; as of June 2026, 1088 academics have registered for access.

Journal of Imaging doi: 10.3390/jimaging12070310

Authors: Xinyu Liu Li Ding Chenlei Lu Wenlong Yang Ping Li

Geosynchronous spaceborne–airborne frequency-modulated continuous-wave bistatic synthetic aperture radar (GEO SA FMCW BiSAR) offers cost-effective and persistent target monitoring. However, both the maneuvers of the receiver during the signal propagation delay and the continuous movements of the radar platforms within the sweep complicate the received echo signal. These factors invalidate the “stop-and-go” assumption, which presumes constant-velocity motion. This paper proposes an echo model that simultaneously considers intra-pulse motion and accelerated motion of the high-maneuvering receiver. The introduction of receiver acceleration leads to nonlinear range terms in the bistatic range history, which will degrade the focusing performance if not properly compensated. Since the acceleration term is a small second-order quantity relative to the time delay, it is approximated by segmenting the aperture and applying the “stop-and-go” assumption within each sub-aperture. After dechirp, the two-dimensional (2-D) spectrum for imaging is derived by applying the principle of stationary phase and determining the azimuth stationary phase point via series reversion. Finally, imaging is achieved by azimuth compression, range cell migration correction, and secondary range compression. Simulation results demonstrate that the proposed algorithm achieves well-focused images while maintaining computational efficiency.

Journal of Imaging doi: 10.3390/jimaging12070309

Authors: Andrzej Burghardt Piotr Garbacz Magdalena Muszyńska

Hybrid imaging methods are emerging as one of the most dynamically evolving research areas in industrial inspection systems. This paper presents a literature review covering relevant scientific publications and official reports on the use of multimodal approaches in quality inspection and NDT systems. Hybrid imaging involves combining two or more imaging techniques to enhance the detection, characterization, and interpretation of features in inspected objects. The paper describes the physical foundations of vision-based inspection systems, including the interaction of optical radiation with matter. It also introduces a classification of optical methods and discusses the role of image fusion in multimodal data processing, with particular emphasis on high-speed quality control systems. The review outlines the current capabilities, limitations, and industrial applications of hybrid imaging, as well as future research directions, including integration with real-time systems and the use of artificial intelligence for automated defect interpretation.

Journal of Imaging doi: 10.3390/jimaging12070308

Authors: Xinyuan Jia Yanling Li Yihui Wang

Image classification interpretability remains a fundamental challenge in the field of computer vision. Despite the remarkable improvements achieved by deep neural networks in classification accuracy, their decision-making processes are often opaque, which limits their applicability in high-stakes scenarios requiring reliability and transparency. Prototype-based methods, such as PIP-Net, address this issue by establishing explicit correspondences between input images and semantic prototypes, thereby enabling an intuitive, evidence-based reasoning paradigm. However, these approaches still suffer from insufficient global context modeling and underutilization of structural relationships among prototypes. To address these limitations, this paper proposes an interpretable image classification model termed PIP-PACA, which is built upon a prototype-aware clustering attention mechanism. In contrast to conventional Transformer architectures based on self-attention, the proposed PACA module introduces a set of learnable cluster centers to project feature representations into a prototype space. Global information is then captured via a bidirectional attention mechanism between features and prototypes. This design is inherently aligned with the principles of prototype learning while reducing the computational complexity from quadratic to linear. Furthermore, a normalization operation is incorporated during the feature extraction stage to enhance the stability of feature distributions and improve the reliability of prototype matching. Extensive experimental results demonstrate that the proposed method not only preserves the interpretability of the original framework but also achieves notable improvements in classification accuracy, sparsity, and prototype purity. These findings validate the effectiveness and superiority of the clustering-based attention mechanism within the prototype learning paradigm.

Journal of Imaging doi: 10.3390/jimaging12070307

Authors: Rohan A. Phadke Samer G. Salman Zane G. Salman Sai M. Yedupati Joshua Ong Alireza Tavakkoli Sainyam Galhotra Ajay Tripuraneni James Rizkalla

Pediatric wrist fractures are among the most prevalent musculoskeletal injuries in children. Fracture subtype, including buckle/torus, greenstick, and Salter–Harris physeal injuries, directly influences management and prognosis. Subspecialty radiographic expertise required for subtype classification is not universally available in emergency or resource-limited settings. Deep learning (DL) offers an automated approach to fracture subtype recognition from plain radiographs. This pilot study evaluated convolutional neural network (CNN)-based five-class pediatric wrist fracture classification using the GRAZPEDWRI-DX dataset.A total of 940 pediatric wrist radiographs from GRAZPEDWRI-DX (figshare ID 14825193) were labeled using Arbeitsgemeinschaft fur Osteosynthesefragen (AO) pediatric codes into five classes: no fracture, buckle/torus, greenstick, Salter–Harris physeal fracture, and other fracture. Contrast-limited adaptive histogram equalization (CLAHE) and letterbox resizing to 224 × 224 pixels were applied. Patient-level stratified splits (70/15/15%) prevented data leakage. Three ImageNet-pretrained architectures (DenseNet-169, ResNet-50, and EfficientNet-B4) underwent two-phase transfer learning. Performance was assessed by balanced accuracy, macro F1, macro area under the receiver operating characteristic curve (AUROC), and Cohen’s kappa.DenseNet-169 achieved the highest balanced accuracy (0.371; 95% confidence interval [CI]: 0.289–0.448), macro F1 (0.334; 95% CI: 0.251–0.416), and macro AUROC (0.669), with Cohen’s kappa of 0.269 on the held-out test set (n = 139) under initial five-epoch pilot training conditions. All three networks exceeded a majority-class (no-information) baseline (balanced accuracy 0.20). Extending training to 50 epochs (approximately 2100 mini-batch iterations) with GPU acceleration substantially improved DenseNet-169 to a balanced accuracy of 0.532 (95% CI: 0.451–0.614), macro F1 of 0.516, and macro AUROC of 0.815, with statistically significant pairwise architecture differences (McNemar p < 0.01); per-class sensitivity was highest for no-fracture detection (0.969) and lowest for buckle/torus fractures (0.393). Gradient-weighted class activation mapping (Grad-CAM) confirmed anatomically coherent model saliency at the distal radial metaphysis and physeal plate.DenseNet-169 achieved the best five-class classification performance among evaluated architectures under pilot training conditions, and extended training substantially improved accuracy, although classification accuracy remained below clinically usable thresholds. These results establish a reproducible, patient-stratified DL pipeline and a benchmark for full-dataset training and future methodological development, rather than a clinically deployable tool.

Journal of Imaging doi: 10.3390/jimaging12070306

Authors: Tian Xia Jianhua Li Liping Sun

Accurate polyp segmentation in colonoscopy supports early detection of colorectal cancer, but compact models under a student-only inference budget tend to lose boundary fidelity. BDKD-Net is a 3.72M-parameter compact student trained under a composite knowledge-distillation loss that combines a response signal, a boundary-probability signal restricted to a static teacher-derived boundary band, and an auxiliary detail-feature alignment; the teacher is used only during training and discarded at inference. In a uniform five-seed re-run on a locked Kvasir-SEG and CVC-ClinicDB development split, the same-architecture scratch student reaches Dev Dice 0.9210 ± 0.0029 and boundary F1 at 3-pixel tolerance 0.7726 ± 0.0057, while full BDKD-Net reaches 0.9300 ± 0.0029 and 0.8008 ± 0.0094. Boundary-probability distillation is the load-bearing signal: it has the highest mean Dev BF1t3 among the single KD signals and carries the boundary gain at no inference cost, with the full model reaching four-external Dice 0.8225 ± 0.0071 at 1.76 GFLOPs and 140.9 FPS. Among the directly reproduced baselines, the higher-Dice Polyp-PVT (0.8417 ± 0.0091) needs 6.8× the parameters and 5.7× the compute at roughly half the frame rate. BDKD-Net thus delivers a compact, boundary-faithful student that keeps most of the accuracy of much larger models at a fraction of their inference cost.

Journal of Imaging doi: 10.3390/jimaging12070305

Authors: Yuting Han Shuyu Li Yunfei Tan

To address the high storage overhead and inadequate spatial geometric representation associated with raw point cloud data in multi-pedestrian trajectory prediction, a compressed point cloud-based and attention-enhanced trajectory prediction method (CPCAE) is proposed in the paper. First, for input raw point cloud, a lossy compression module is designed, which improves the Depoco framework by introducing a multi-feature extraction component and employing a coordinate decomposition strategy to optimize compression quality and spatial representation. For input video frames of pedestrians, spatial features are extracted using a 2D convolutional network, and dynamic interactions among pedestrians are captured by a Transformer-based encoder. Then, both spatial attention and modal attention mechanisms are incorporated to dynamically balance the contributions of two modal features and precisely identify key regions and positions. Experimental results evaluate the proposed framework from the perspectives of point cloud compression and downstream trajectory prediction. The results demonstrate that compressed point cloud representations can support competitive trajectory prediction performance in CPCAE.

Journal of Imaging doi: 10.3390/jimaging12070304

Authors: Xue Zhao Zhuoyue Hu Zhengqin Xu

Vignetting introduces spatial radiometric nonuniformity into remote sensing images and degrades subsequent radiometric analysis, image interpretation, and calibration-related applications. To address this problem, this paper proposes a vignetting correction method based on low-rank modeling and polynomial fitting. The method constructs a multi-frame data matrix in the logarithmic domain, extracts the shared vignette component through rank-1 low-rank modeling, and further recovers a smooth vignette field through polynomial fitting. Experiments were conducted using real remote sensing images, simulated vignetted images, and star images. Among the three ablation variants, the proposed full method achieved the best performance, with MAE, MAD, Center-MAE, and Edge-MAE values of 0.48%, 3.65%, 0.14%, and 0.52%, respectively. Compared with the low-rank-only method, these metrics were reduced by 23.8%, 32.8%, 71.4%, and 20.0%, respectively. An additional all-frame comparison across 28 dataset settings showed that the proposed rank-1 model achieved mean accuracy comparable to nuclear-norm-based standard RPCA, while exhibiting lower cross-dataset variability in MAE, MAD, and Edge-MAE. For star images, the method reduced image-plane nonuniformity from 1.39–1.92% to 0.59–0.80% while preserving background-subtracted stellar DN values. These results demonstrate that the proposed method provides physically interpretable and stable vignetting correction while maintaining radiometric consistency.

Journal of Imaging doi: 10.3390/jimaging12070303

Authors: Fatma Küçük

Hyperspectral anomaly detection application is essential in numerous different remote sensing applications, where the detection of rare and unknown targets in complex scenes is needed. The proposed anomaly detection method in this study is called L-PCAD (Laplacian PCA-based anomaly detection), which plans to combine a Linearized Alternating Direction method (LADM)-based subspace recovery algorithm with a modified RX detector to improve detection accuracy and stability. It starts with an LADM-based approach as a preprocessing stage to get a matrix with rich information relating to anomalies. The resulting low-rank background matrix is subsequently utilized as a guide for the anomaly detection process. In order to enhance the RX detector, the covariance estimation is reformulated using a graph Laplacian constructed from the low-rank background matrix. Instead of directly using the empirical covariance matrix, a normalized Laplacian is computed and subsequently transformed via principal component analysis (PCA) to obtain a stable diagonal representation. This PCA-regularized Laplacian replaces the conventional covariance matrix in the RX formulation while preserving the local spatial structure. The extensive testing of different hyperspectral datasets shows that the proposed approach provides better overall results for detection performance as compared to other hyperspectral anomaly detectors that are the state of the art.

Journal of Imaging doi: 10.3390/jimaging12070302

Authors: Xiaolong Hu Xueyan Liu Junji Jiang Ziqi Hao Lishan Qiao

Medical image segmentation is critical for clinical diagnosis, yet existing methods face a persistent trade-off: CNN-based approaches are constrained by local receptive fields, while Transformer-based methods suffer from semantic dilution when modeling global context. To address these limitations, we propose SCAGC-UNet, a region-aware graph convolutional network that bridges local detail extraction and global dependency modeling through structured region-level reasoning. The architecture features a dual-layer residual encoder for hierarchical feature extraction and a Spatial-Channel Graph Convolution (SC-GCN) module at the bottleneck, which simultaneously captures inter-region spatial topology and intra-region channel semantics via dual-branch graph inference. Feature refinement in the decoder is further enhanced by Context-Corrected Modules and Backward-Aided Modules to reduce the semantic gap across skip connections. We validate SCAGC-UNet on three public benchmarks covering distinct imaging challenges. On Kvasir-SEG, the model achieves a Dice score of 92.28% and MIOU of 92.41%, surpassing the strongest CNN-based baseline CCBANet by 0.73% in DSC and outperforming TransUNet by 11.76% in DSC. On BUSI, it attains an IOU of 78.10% and MIOU of 87.68%, outperforming UNet by 2.82% in IOU and TransUNet by 6.91% in DSC. On COVID-19 CT, it achieves a DSC of 82.51%, surpassing UNet by 4.99% and TransUNet by 7.47%, demonstrating robust performance on irregular lesion morphologies. These results confirm that SCAGC-UNet achieves consistent and robust segmentation performance across three public benchmark datasets spanning distinct imaging modalities, suggesting its potential clinical relevance.

Journal of Imaging doi: 10.3390/jimaging12070301

Authors: Evans Ohemeng Andrew Donkor Ijeoma Chinedum Anyitey-Kokor Obed Kojo Otoo Theophilus N. Akudjedu William Kwadwo Antwi Yaw Amo Wiafe

While several dosing protocols for iodinated contrast media (ICM) exist, a consensus strategy for optimising the physical imaging signal in abdominal CT is lacking. This systematic review and meta-analysis evaluated the performance of individualised dosing protocols, specifically focusing on technical signal optimisation, clinical safety, potential material savings, and environmental sustainability. Electronic databases (Cochrane, Embase, Medline) were searched up to January 2026. Systematic synthesis of 23 studies (11,680 participants) compared protocols based on lean body weight (LBW), total body weight (TBW), fixed volume (FV) and software-assisted dosing. Meta-analyses assessed volume optimisation and hepatic enhancement, with evidence certainty evaluated via the GRADE framework. TBW-based dosing significantly reduced contrast volume by −8.74 mL compared to FV protocols (p = 0.02), while the −4.04 mL reduction in LBW versus TBW groups represented a non-significant trend (p = 0.11); however, a sensitivity analysis revealed a significant effect (−5.41 mL, 95% [CI: −10.43, −0.39]; p = 0.03). Pooled hepatic enhancement showed no statistically significant differences for LBW vs. TBW (−1.36 HU, p = 0.43) or FV vs. TBW (−2.74 HU, p = 0.13). Individualised ICM dosing, particularly LBW-based, may potentially offer a foundational strategy for greener and material savings in clinical radiology by minimising population-level iodine load. Despite modest individual volume reductions, these protocols may potentially facilitate standardised imaging enhancement, though higher-quality randomised trials are required to confirm safety and economic benefits.

Journal of Imaging doi: 10.3390/jimaging12070300

Authors: Alexandru Mitoi Raluca Mititelu Cosmin Medar Vlad Octavian Bolocan Constantin Ciprian Ioan-Nicolae Mateș

Radiomic feature stability is a necessary condition for clinical translation, yet the impact of inter-observer segmentation variability remains insufficiently characterized for [18F]FDG PET/CT in breast carcinoma. We evaluated the inter-observer reproducibility of 107 original radiomic features extracted from [18F]FDG PET/CT images of 42 patients with biopsy-proven, treatment-naive primary breast carcinoma, using an IBSI-aligned PyRadiomics workflow. Two nuclear medicine physicians independently segmented each tumor using semi-automatic Otsu thresholding to generate an initial tumor mask, followed by manual correction. Reproducibility was quantified using ICC(A,1) with bootstrap-derived 95% confidence intervals. A two-stage reproducibility and redundancy-based feature reduction strategy, combining an ICC threshold with Spearman correlation-based redundancy removal, was applied across nine threshold combinations, and features were classified into three pre-specified stability categories. The segmentation agreement was good, with a mean Dice coefficient of 0.847. Most features showed excellent reproducibility (81/107, 75.7% with ICC ≥ 0.90; median ICC 0.972), whereas shape features based on maximum lesion extension showed poor reproducibility (ICC 0.10–0.25). The reduction strategy resulted in 19 stable non-redundant features, with eight retained across all threshold combinations; 79 features (73.8%) met high-stability criteria. These results and the proposed stability classification framework provide a methodological basis for future predictive PET radiomics studies in breast carcinoma.

Journal of Imaging doi: 10.3390/jimaging12070299

Authors: Jiahao Guo Tao Chen Jiaxi Hu Yuanhong Zhou

Accurate medical image segmentation remains challenging when lesions have large-scale variation, weak boundaries, and strong background interference. Vision Mamba provides efficient long-range modeling, but current Mamba-based U-shaped networks are still limited by weak local multi-scale representation and coarse skip fusion. This study proposes MSCF-Net, a Vision Mamba segmentation network for dermoscopic and endoscopic images. The network is built on VM-UNet and introduces two modules. The Multi-Scale Context Bridging (MSCB) module enriches bottleneck features with local, dilated, and global context. The Cross-Layer Adaptive Fusion (CLAF) module recalibrates encoder–decoder features in channel and spatial dimensions, reducing noisy shallow feature transmission. A structure loss is used to improve region completeness and boundary quality. Experiments on ISIC 2017, ISIC 2018, and CVC-ClinicDB show Dice scores of 90.62%, 90.82%, and 91.72%, and mIoU values of 82.02%, 82.31%, and 84.56%, respectively. Compared with representative baselines evaluated in our experiments, MSCF-Net achieves competitive segmentation performance under the adopted benchmark protocol. Ablation, qualitative, and spatial response analyses further indicate that MSCB improves scale-aware representation, while CLAF helps the decoder focus on lesion-related cues. The results suggest that MSCF-Net provides a favorable accuracy–efficiency trade-off for medical image segmentation.

Journal of Imaging doi: 10.3390/jimaging12070298

Authors: Syed Baqir Hussain Shah Musfarah Wajid Syed Adil Hussain Shah Silvia Godio Karim Kassem Gohar Bano Zaidi Shahzad Ahmad Qureshi Syed Taimoor Hussain Shah Marco Agostino Deriu

Accurate automated interpretation of upper-extremity musculoskeletal radiographs remains challenging because fracture appearance varies across anatomical regions and can be subtle under class imbalance. This study proposes a two-stage deep learning framework for MURA-based X-ray analysis, aiming to improve body-part recognition and body-part-wise abnormality detection. Multiple architectures were first compared for seven-class body-part classification, after which the selected hybrid Xception-Swin model was fine-tuned for abnormality detection within each anatomical subset. The framework combines Xception-derived local structural features with Swin Transformer contextual features using attention-based fusion, and performance was evaluated using accuracy, F1-score, AUC-ROC, Cohen’s kappa, calibration, component-level ablation, post hoc explainability, and zero-shot FracAtlas validation. For body-part classification, the model achieved accuracy = 0.9643, macro F1 = 0.9574, AUC-ROC = 0.9963, and kappa = 0.9579. For abnormality detection, accuracy ranged from 0.7289 to 0.8538, F1 from 0.7191 to 0.8508, AUC from 0.7693 to 0.9080, and kappa from 0.4449 to 0.7071. Ablation on hand and humerus radiographs showed the highest macro F1 with Hybrid Attention, while FracAtlas validation yielded AUC = 0.8247 and kappa = 0.5812. The results support complementary CNN-Transformer fusion and indicate preliminary cross-dataset generalizability. Implementation resources are available at Zenodo.

Journal of Imaging doi: 10.3390/jimaging12070297

Authors: Diego Martí-Pérez Valery Naranjo Adrián Colomer

Cell segmentation plays a key role in a wide range of biomedical imaging applications, from single-cell analysis to pathology assessment. While classical deep learning architectures such as U-Net, StarDist, and HoVer-Net have set strong baselines, their reliance on domain-specific training limits generalization across diverse microscopy modalities. The emergence of foundation models, particularly the Segment Anything Model (SAM) and its derivatives, has introduced a paradigm shift toward more universal and adaptable segmentation frameworks. In this review, we summarize key advances in microscopy cell segmentation, highlighting both traditional methods and recent foundation model-based approaches. Beyond surveying the literature, we present an experimental comparison of four representative models—our proposed YOLO-SAM, along with CellSAM, Cellpose-SAM, and StarDist—tested on both fluorescence and brightfield microscopy spanning diverse cell populations and shapes. Our findings illustrate trade-offs between accuracy, robustness, and adaptability, with foundation-based models showing particular promise for cross-domain performance. By combining a comprehensive review with systematic benchmarking, this work provides practical guidance for researchers and outlines current challenges and future opportunities in developing robust, generalizable cell segmentation methods for microscopy.

Journal of Imaging doi: 10.3390/jimaging12070296

Authors: Ivan De Boi Rhys G. Evans Stuti Pathak Thomas De Kerf Marnix Van Soom Sam Van der Jeught Helder Araújo Rudi Penne

Accurate lens distortion correction is important for calibration, registration, image stitching, and 3D reconstruction, especially in low-data device-specific settings where disposable or specialised cameras cannot provide large calibration datasets. We address distortion correction for cameras with highly irregular or non-stationary distortion fields, where fixed polynomial models and generic learning-based rectification methods can struggle. We propose a framework based on Deep Gaussian Processes (DGPs) to model the non-linear mapping required for undistortion. The key motivation is that conventional single-layer GPs with stationary kernels must use one global notion of smoothness, whereas DGPs can represent spatially varying behaviour through composed latent mappings while preserving per-pixel predictive uncertainty. This uncertainty can be used to identify or downweight unreliable corrected regions in downstream tasks. We evaluate the method on three real camera datasets with increasing distortion complexity. The full structured acquisitions contain 512 horizontal and 512 vertical line images per camera. These are not thousands of natural calibration images, but they yield up to 29,532, 11,311, and 31,686 detected intersection correspondences for the RPI, Theta, and Pillcam datasets, respectively. This distinction is important for cameras where acquiring many independent images is impractical. The results are assessed using qualitative rectification, uncertainty maps, normalised collinearity errors, and total training time. Polynomial calibration remains strongest for the regular radial RPI distortion, while DGP and DGP2 models show lower normalised collinearity-error distributions than the standard GP and lightweight MLP baselines on the more distorted Theta and Pillcam datasets. For the full datasets, total DGP/DGP2 training times ranged from 2383.50 s to 10092.50 s, reflecting the additional computational cost of probabilistic non-stationary modelling.

Journal of Imaging doi: 10.3390/jimaging12070295

Authors: Héctor Guillen-Bonilla José Trinidad Guillen-Bonilla Maricela Jiménez-Rodríguez Alex Guillen-Bonilla Jorge Aguilar-Santiago Lucía Ivonne Juárez-Amador Antonio Casillas Zamora

In this paper, an RGB image with S is separated by its channels, obtaining an image in each color channel SR, SG and SB. The Vectorial Image Representation on the Texture Space (VIR-TS) transform is calculated for each channel; ergo, each image is represented with a given vector, SR→ C→R, SG→C→G and SB→C→B. Employing the C→R, C→G, and C→B vectors in a multi-class classifier, a database of RGB images was identified with the aim of verifying the classification efficiency of the VIR-TS transform. Based on the experimental results, the VIR-TS technique presents high efficiency when the noise is not added to the class and when the signal-to-noise ratio is high. For both instances, the efficiency presented is 100%. Nonetheless, if the class noise is high, the efficiency diminishes from 100%, to 95%, to 90% until it decreases to 10%. Based on the results obtained, the VIR-TS transform can be efficiently applied for the development of security systems and access control and can also be implemented in computer vision systems for medical diagnosis, drones, etc.

Journal of Imaging doi: 10.3390/jimaging12070294

Authors: Zixuan Li Jiaxin Liu Hongwei Wang Zhaoming Liu Feng Zhang Ning Bai Jing Hou Yongliang Yang Long Cui

Aero-engine blades operate under extreme conditions involving high temperature, pressure, rotational speed, and cyclic loads, making them susceptible to surface defects such as micro-cracks. Due to their small scale, weak edges, low contrast, and elongated morphology, micro-cracks are easily affected by metallic reflections, uneven illumination, and complex background textures in borescope images, resulting in high missed-detection rates for conventional detection methods. To address these challenges, this study proposes an improved YOLO11-based framework for aero-engine blade micro-crack detection. The proposed method introduces P1/P2 shallow high-resolution detection branches to enhance the perception of fine crack edges and textures, incorporates Focal Loss to alleviate foreground–background imbalance, applies object-level Tversky Loss to strengthen false-negative constraints, and adopts a hard mining strategy to improve learning for difficult crack samples. Experiments conducted on a real aero-engine borescope image dataset demonstrate that the proposed model achieves a Precision of 0.9981, Recall of 0.9606, F1-score of 0.9790, mAP50 of 0.9781, and mAP50-95 of 0.6938 on an independent test set. Compared with the YOLO11 baseline, the proposed method significantly improves crack detection accuracy, localization quality, and robustness in complex borescope inspection scenarios.

Journal of Imaging doi: 10.3390/jimaging12070293

Authors: Shiqi Zhang Peng Li

Class imbalance remains a critical challenge in image classification, where underrepresented classes often receive insufficient training attention and exhibit poor recognition performance. In this study, we propose a hybrid framework that combines weighted sampling with frequency-aware spatial attention (WSFSA) to address class imbalance at both the data and feature levels. The weighted sampler improves the training exposure of minority classes, while the frequency-aware spatial attention module incorporates class-frequency information into spatial attention to enhance discriminative feature responses for underrepresented classes. We evaluate the proposed method on four MedMNIST benchmarks, DermaMNIST, BloodMNIST, OrganCMNIST, and DermaMNIST-224, using a ResNet-18 backbone. Results show that WSFSA provides the clearest benefit on the severely imbalanced DermaMNIST and DermaMNIST-224 datasets, performing comparably to the strongest baseline methods while showing particular benefits under severe class imbalance. On OrganCMNIST, WSFSA provides moderate gains, while on BloodMNIST, where the imbalance effect is weaker, all methods perform similarly. Per-class analysis further shows that WSFSA improves sensitivity for several minority or difficult classes while maintaining high specificity across most classes. These findings suggest that combining sampling-level and feature-level rebalancing is a practical strategy for improving class-balanced recognition, particularly under severe class imbalance.

Journal of Imaging doi: 10.3390/jimaging12070292

Authors: Xiansong Gu Qiang Fu Zhuang Liu Guan Wang Hairui Wang Chao Wang Tianshu Wang Yingchao Li Huilin Jiang

The integration requirements of laser ranging, imaging, and communication functions in space target detection have placed higher demands on system performance. This paper takes a modularly designed integrated laser ranging, imaging, and communication system as an example and proposes a light source integration scheme based on fiber phased array beam splitting–coupling technology, effectively enhancing the system’s integration level and compactness. The system employs a Cassegrain optical system and beam splitting structure to achieve functional integration of laser communication, ranging, and polarization imaging. Ground experiments were conducted to evaluate the functional feasibility of the proposed integrated architecture. The visible light polarization imaging experiments at kilometer-level distances demonstrate that polarization-derived information can improve target–background separability under haze and low-contrast conditions. The UAV-based dynamic ranging experiment verifies that the system can acquire, track, and range a moving cooperative target under the tested field conditions, with the measured results being consistent with the designed meter-level ranging requirement. In addition, a 1 km coherent free-space laser communication experiment achieved 20 Gbps QPSK signal transmission with a bit error rate on the order of 10−7. These results provide experimental support and design references for integrated optoelectronic terminals used in space target observation, space debris monitoring, and related long-distance sensing and communication applications.

Journal of Imaging doi: 10.3390/jimaging12070291

Authors: Zhenjun Yu Lin Sun Kai Wang Fengxiang Jin

Knowledge distillation is a cornerstone of image anomaly detection for amplifying subtle defects via teacher–student discrepancy, yet existing methods rely on feature alignment loss that causes reconstruction error confusion and degrades accuracy. To address this critical limitation, this study proposes an abnormal discrepancy-guided knowledge distillation method (DiffKD) that differentially guides student feature reconstruction through channel-level discrepancy masks, leveraging normal features as supervisory signals and abnormal discrepancy features as constraints to enhance anomaly detection performance. The approach integrates a knowledge distillation network for feature reconstruction with a segmentation network for anomaly localization, while utilizing prior anomaly samples and synthetic anomaly samples to provide real-time training data of anomalous samples. Extensive evaluations on the SUT-Crack and MVTec AD benchmarks validate the effectiveness and generalizability of our approach. On MVTec AD, it achieves 80.7% average precision (AP) and 81.9% instance-level average precision (IAP), showing competitive performance against the representative methods evaluated under the same protocol. These results not only demonstrate significant improvements in IAD accuracy but also highlight its promise for enabling real-time, automated anomaly detection in practical applications.

Journal of Imaging doi: 10.3390/jimaging12070290

Authors: Lixin Zhou Yuanyuan Yang Yunfeng Yang

Transformer, with its unique self-attention mechanism, naturally excels in modeling global features. Convolutional Neural Networks (CNNs), on the other hand, leverage strong spatial inductive biases to effectively capture local features with fewer parameters. In 3D brain tumor segmentation, both local and global features are critical. Moreover, balancing high accuracy with computational cost in 3D segmentation models remains a key challenge. To address this, we propose TransLiteUNet, a lightweight 3D solution that combines CNN and Transformer architectures for accurate brain tumor segmentation without pretraining. To enhance parameter efficiency, we introduce a 3D axial depthwise separable convolution residual structure (3DRes-ADS block) and a lightweight LiteViT module, which improves global feature modeling at a lower computational cost. Specifically, TransLiteUNet (0.43 M parameters, 14.98 GFLOPs) and its simplified version, TransLiteUNet-S (0.31 M parameters, 7.68 G FLOPs), offer significantly lower model complexity compared to current state-of-the-art models. Tested on two publicly available datasets, our models outperform leading models under identical conditions. The parameter and computational costs are reduced by orders of magnitude, with optimized inference and training costs.

Journal of Imaging doi: 10.3390/jimaging12070289

Authors: Shuai Yan Shirong Feng Zhicheng Wei

Semantic segmentation in foggy scenes is crucial for autonomous driving systems, yet acquiring annotated real-world foggy data is highly costly. Existing unsupervised domain adaptation methods typically adopt a self-training strategy, adapting models trained under clear-weather conditions to unlabeled foggy target domains and constructing supervision signals via pseudo-labels. However, these methods mainly focus on improving the reliability of target-domain supervision while paying insufficient attention to class confusion caused by the degradation of class discriminability. In fact, the performance degradation of self-training in foggy scenarios is not caused by a single factor, but is jointly affected by unreliable supervision signals and reduced class discriminability. To address these issues, this paper proposes a reliable pseudo-labeling and confusion calibration framework for foggy-scene semantic segmentation, termed RPCC. Specifically, dynamic energy-guided pseudo-labeling (DEPL) models the reliability of target-domain predictions using energy scores, thereby improving the reliability of target-domain supervision signals. Furthermore, the reliable-region class confusion calibration (RCC) module models and calibrates semantic class relationships in target-domain predictions based on reliable pseudo-label supervision, thereby suppressing class confusion and enhancing semantic boundary clarity. Experimental results demonstrate that RPCC outperforms existing methods on multiple real-world foggy-scene datasets and shows favorable generalization to other adverse weather conditions.

Journal of Imaging doi: 10.3390/jimaging12070288

Authors: Behnam Kiani Kalejahi Sajid Khan Mohammad Javad Rajabi

Accurate, automated delineation of adult diffuse gliomas from multi-parametric magnetic resonance imaging (mpMRI) is central to quantitative neuro-oncology. Volumetric 3D networks dominate the BraTS leaderboard but require expensive GPUs, long training cycles, and provide diminishing returns relative to their compute budget. Slice-wise 2D models, by contrast, discard inter-slice context that is informative for thin tumor rims and small enhancing foci. We introduce GDNet, a 2.5D multimodal MRI segmentation framework for adult glioma evaluated on the BraTS 2024 cohort. GDNet consumes a stack of three adjacent axial slices from the four standard BraTS modalities (T1, T1ce, T2, FLAIR) as a 12-channel input to a compact U-shaped encoder–decoder with Group Normalization and predicts whole tumor (WT), tumor core (TC), and enhancing tumor (ET) masks for the central slice. The training pipeline pairs the 2.5D backbone with: (i) Exponential Moving Average (EMA) of model weights with decay 0.999, (ii) mixed tumor-aware slice sampling (p_tumor = 0.50), (iii) a compound Cross-Entropy + Soft-Dice loss, and (iv) AdamW with warm-up plus cosine annealing under Automatic Mixed Precision. We performed a systematic, step-by-step ablation covering a 2D baseline, EMA + mixed sampling, tumor-centered crop fine-tuning, a GDNet-inspired architectural integration, a region-aware loss, 3-slice and 5-slice 2.5D inputs, and connected-component post-processing, and we report multi-seed results to quantify reproducibility. On the held-out BraTS 2024 test partition, the final 3-slice 2.5D GDNet achieved positive-only Dice scores of 0.791 ± 0.000 (WT), 0.736 ± 0.003 (TC), 0.654 ± 0.004 (ET), and a mean foreground positive-only Dice of 0.820 ± 0.000 across seeds; the all-slice mean foreground Dice exceeded 0.927 ± 0.000. Validation positive-only scores were 0.805 ± 0.002 (WT), 0.757 ± 0.004 (TC), 0.683 ± 0.009 (ET). The inter-seed standard deviation was small for every region (≤0.01 Dice points), indicating low inter-seed variance across the two seeds evaluated; with only two seeds, we regard this as preliminary evidence of training stability rather than a strong reproducibility claim. The ablation isolated EMA + mixed tumor sampling and the 2.5D context window as the dominant sources of improvement; notably, a GDNet-style architectural integration with a region-aware loss did not outperform the simpler 2.5D U-Net on positive-only WT/TC/ET, and light post-processing improved only all-slice Dice. A failure-mode audit found that the residual catastrophic predictions are concentrated on a small minority of diffuse, infiltrative tumors with mass effect. Conclusions: Carefully engineered training strategies, tumor-aware sampling, EMA stabilization, and a modest 2.5D context window recover a substantial fraction of the accuracy of much heavier 3D networks at a fraction of the compute, are reproducible across seeds, and outperform a heavier GDNet-inspired architectural variant on the same data. GDNet is therefore a practical and, pending external validation, potentially clinically deployable framework for multimodal glioma segmentation on workstation-class GPU hardware.

Journal of Imaging doi: 10.3390/jimaging12070287

Authors: Andrea Lastrucci Nicola Iosca Edoardo Cavigli Diletta Cozzi Angelo Barra Yannick Wandael Cosimo Nardi Renzo Ricci Vittorio Miele Daniele Giansanti

Background: Lung cancer remains the leading cause of cancer-related mortality worldwide, and early diagnosis and accurate disease stratification are still major clinical challenges. Radiomics has emerged as a quantitative imaging approach that extracts high-dimensional features from radiological imaging, with applications in diagnosis, prognosis, radio genomics, and assessment of treatment response. However, its clinical translation is still limited by methodological heterogeneity and a lack of standardization. Aim: This narrative review synthesizes evidence from systematic reviews and meta-analyses on radiomics in thoracic imaging for lung cancer, focusing on clinical applications, methodological limitations, and translational challenges. Methods: A structured search was conducted in PubMed and Scopus using predefined keywords related to radiomics, lung cancer, and imaging modalities. Only peer-reviewed systematic reviews and meta-analyses published in English were included. In total, 27 studies were selected and synthesized using a structured narrative approach guided by the ANDJ checklist. A differential integrative framework was adopted to connect evidence from systematic reviews and meta-analyses with primary empirical studies and policy documents through an intermediate layer of translational recommendations, ensuring a multi-level and interpretation-driven synthesis. Results: Radiomics demonstrated consistent potential across multiple clinical domains, including lesion classification, histological differentiation, molecular profiling, prognostic stratification, and prediction of treatment response. Machine learning and deep learning approaches frequently improved predictive performance. However, key limitations were identified, including heterogeneity in imaging protocols, lack of external validation, small single-centre datasets, and limited reproducibility of radiomic features. Conclusions: Radiomics in lung cancer imaging shows strong clinical potential but remains constrained by methodological and translational barriers. Future progress will depend on standardization, external validation, multimodal data integration, and improved interpretability, alongside alignment with regulatory and clinical implementation frameworks.

Journal of Imaging doi: 10.3390/jimaging12070286

Authors: Khaldoon Alhusari Salam Dhou

Breast cancer is the most commonly diagnosed cancer in women, with early detection playing a critical role in clinical outcomes. Mammography remains the standard screening modality, producing X-ray images used to assess mammographic density, a key indicator of the proportion of fibroglandular tissue within the breast. The Breast Imaging-Reporting and Data System (BI-RADS) classification system is widely used to report density across four qualitative categories. High density can obscure malignancies and is independently associated with elevated breast cancer risk. Manual interpretation of mammographic density is prone to subjectivity and inter-observer variability, and supervised learning-based estimation methods trained on subjective labels may reflect this inherent subjectivity. This work proposes an unsupervised framework for quantitative breast density estimation that requires no labeled data in its core pipeline. Expert labels are used exclusively to calibrate post hoc discretization thresholds for binary classification, enabling comparison with supervised methods in the literature. The main contributions include: (i) an adaptive Region of Interest (ROI) extraction algorithm, (ii) a Convolutional Neural Network (CNN) based unsupervised segmentation pipeline tuned for mammographic density separation, (iii) a novel confidence metric for identifying unreliable segmentation outputs, (iv) a label correction mechanism for low-confidence cases, and (v) a confidence-filtered majority voting scheme for per-patient classification. The framework is evaluated on two public datasets, namely DDSM and INbreast, with segmentation performance yielding Silhouette scores exceeding 0.92. Agreement with expert labels reaches 71.43% and 79.28% for DDSM and INbreast, respectively. Image-level clustering quality assessment confirms effective unsupervised labeling, with Silhouette scores averaging 0.57 for DDSM and 0.50 for INbreast. The proposed framework provides a practical and non-subjective model for quantitative breast density estimation, with potential utility as a decision-support tool for radiologists that can be considered in clinical practice after further investigation.

Journal of Imaging doi: 10.3390/jimaging12070285

Authors: Michael Moncher Vincent Traboulsi Florian Kofler Sarah Müller Felix Steinbauer Constantin von See

Osteoporosis may alter mandibular bone structure and peri-implant remodeling, but it remains unclear whether such changes are detectable on dental radiographs using deep learning. This retrospective study evaluated whether osteoporosis can be discriminated in two mandibular regions of interest: peri-implant bone and the mental foramen region. Digital periapical radiographs acquired between November 2012 and October 2024 were analyzed in 51 women, including 25 patients with osteoporosis and 26 non-osteoporotic controls without a documented history or diagnosis of osteoporosis; the osteoporosis group was significantly older than the control group. Two binary classification experiments were performed using patient-level fivefold grouped cross-validation. The peri-implant experiment included 1682 cropped images and used an image-plus-metadata ResNet-18 model incorporating the time interval between implant placement and radiograph acquisition. The mental foramen experiment included 102 cropped images and used an image-only ResNet-18 model. Mean accuracy, F1 score, and area under the receiver operating characteristic curve were 0.613, 0.628, and 0.713 for the peri-implant region of interest (ROI) and 0.701, 0.713, and 0.744 for the mental foramen ROI, respectively. Both experiments showed substantial fold-to-fold variability. These findings suggest that ROI selection influences model behavior, but neither approach yielded sufficiently stable ROI-level classification performance under patient-level grouped validation to support individual patient-level screening claims. Nondiscriminatory AI results should therefore be interpreted as limited evidence under the present experimental conditions rather than as proof of radiographic equivalence.

Journal of Imaging doi: 10.3390/jimaging12070284

Authors: Yaxin Xie Mingyu Zhang Yonghua Han Le Dai Haifeng Fu Lu Xu

In cotton disease detection, the complex farmland environment and the varying scales of disease spots, especially the presence of small-target disease spots, limit the detection accuracy of lightweight models. To address this issue, an improved YOLOv11n detection algorithm is proposed. First, the backbone network is reconstructed using the GhostConv (G-conv) module, which generates redundant feature maps through linear operations, thereby reducing computational complexity. Second, an Adaptive Calibration and Feature Fusion Architecture Head (ACFFA) with prior calibration and cross-scale fusion capabilities is constructed in the detection stage to handle the problem of varying disease spot scales. Furthermore, the Adaptive Scale-aware Wise Intersection over Union (AS-WIoU) loss function, improved from WIoUv3, is introduced to enhance the stability of bounding box regression and improve detection accuracy for low-resolution, small-target lesions. Experimental results show that on the cotton disease dataset constructed based on the Mendeley Data database, the proposed model achieves mAP50 and mAP50-95 of 90.30% and 73.84%, respectively, with precision and recall of 92.33% and 87.68%, and a parameter count of 3.81 M. The algorithm significantly improves detection accuracy while maintaining efficient inference, making it suitable for real-time monitoring tasks on agricultural embedded terminals.

Journal of Imaging doi: 10.3390/jimaging12070283

Authors: Ngoc-Anh A. Nguyen Grace Lee Sarah Sossong Jannika V. Machnik Sarah Pletcher Roberta Schwartz

Background: Operating rooms generate substantial visual data that is rarely captured systematically. Image-based AI (IBAI) systems using computer vision offer a new approach to real-time perioperative workflow monitoring, but evidence of their impact on surgical case volume remains limited. The aim of this study was to evaluate the association between deployment of an IBAI system and monthly surgical case volume in a high-acuity cardiothoracic operating suite, using synthetic control with difference-in-differences estimation. Methods: We deployed an IBAI system with wall-mounted cameras and a YOLO-based (You Only Look Once) object detection model coupled with a transformer-based event detector in a 15-room cardiothoracic suite at Houston Methodist Hospital (HMH), the tertiary academic hospital of Houston Methodist health system. The deployment was conducted under an IRB-determined quality improvement framework with patient consent for ambient video capture, defined retention limits, and restricted access to recordings. Over a 16-month period spanning 6 months pre-deployment and 10 months post-deployment, the system monitored 5417 surgical cases and automatically detected additional perioperative events including patient entry, draping, and room turnover. Using a synthetic control methodology, we compared post-deployment outcomes at the intervention site against a weighted combination drawn from a pool of 11 Houston Methodist sites that did not yet implement IBAI (116,098 cases across the comparison sites; 121,515 cases in the full analytic dataset). Results: The synthetic control analysis with difference-in-differences estimation showed a statistically significant increase of approximately 25 cases per month (95% CI 8.3 to 41.0; p < 0.01; Bonferroni-adjusted p < 0.05), corresponding to a 7% increase in monthly case volume relative to baseline. Conclusions: Our findings suggest that IBAI can meaningfully improve OR efficiency and support data-driven perioperative management. Future work should evaluate whether case volume gains generalize across other surgical specialties, assess changes in operational outcomes such as turnover time and first-case on-time starts, and examine clinicians’ perceptions of IBAI.

Journal of Imaging doi: 10.3390/jimaging12070282

Authors: Xiansong Gu Chao Wang Huilin Jiang Boshi Wang

We propose a dual-aperture, simultaneous-polarization super-resolution imaging system that combines a total internal reflection optical architecture with a digital micromirror device (DMD) for broadband, high-resolution imaging. The system captures multiple polarization states simultaneously with a single detector and offers a compact, lightweight design. Using reflective Wassermann–Wolf differential equations and Seidel aberration theory, we establish astigmatism-correction boundary conditions and apply iterative optimization to jointly correct spherical aberration, coma, astigmatism, and distortion. Because distortion critically affects super-resolution reconstruction by causing mirror–pixel misregistration, we further introduce a custom merit function to tightly constrain chief-ray positions for each sub-aperture and field point on intermediate and final image planes, effectively suppressing distortion. The final design achieves F/2.5, grid distortion below ±0.5%, and near-diffraction-limited performance in all polarization channels. Tolerance analysis of the four sub-apertures confirms that imaging requirements are satisfied, demonstrating robust high-resolution polarization imaging across multiple polarization states.

Journal of Imaging doi: 10.3390/jimaging12070281

Authors: Miguel-Angel Gil-Rios Nivia Escalante-Garcia Juan C. Valdiviezo-Navarro Paola Andrea Mejia-Zuluaga León Dozal Ivan Cruz-Aceves

This paper proposes a novel method for automatically designing lightweight Convolutional Neural Network (CNN) architectures. (1) Background: Automated remote sensing for vegetation monitoring faces challenges from structural complexity and cluttered backgrounds. For detecting parasitic Phoradendron velutinum infestations, existing vision frameworks rely on handcrafted, overparameterized CNNs, limiting deployment on localized edge computing platforms. (2) Methods: To address this efficiency-accuracy trade-off, a two-phase hybrid multi-objective Neural Architecture Search (NAS) strategy is implemented. First, the Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) minimizes classification error and the number of trainable parameters. Second, an Iterated Local Search (ILS) metaheuristic refines promising non-dominated solutions. The approach was evaluated using cost-effective aerial RGB imagery, processing a balanced dataset of 5000 patches (64×64 pixels) under a rigorous three-way data partition to prevent data leakage. (3) Results: The discovered 10-layer CNN topology achieved high feature-extraction efficiency. On the unseen testing set, the model yielded an Accuracy and F1-Score of 0.979, a Precision of 0.982, a Recall of 0.976, and a Jaccard Index of 0.958, outperforming the compared models. Operating with only 2040 trainable parameters, the optimized architecture establishes a highly viable paradigm for real-time digital image processing on hardware-constrained monitoring devices.

Journal of Imaging doi: 10.3390/jimaging12070280

Authors: Madalina Carbureanu Florin-Stefan Zamfir

As calibration errors have a direct impact on epipolar consistency, rectification accuracy, and metric 3D reconstruction performance, stereo camera calibration is a fundamental requirement for high-accuracy 3D modeling and reliable digital twin data acquisition. Because current calibration workflows (based on pairwise calibration methods) lack systematic data-quality checks mechanisms, there is a clear need for more robust data selection strategies. The novelty of the approach consists in the development of a new outlier-aware stereo calibration algorithm (OutAw) that introduces a unified multi-stage approach that integrates hard geometric selection, candidate subset generation, multi-criterion ranking, bootstrap stability analysis, and triangulation assessment into a comprehensive and systematic calibration framework. Unlike conventional approaches, OutAw (through its mechanism of detecting and rejecting inconsistent pairs) redefines the calibration strategy from arbitrary to criterion-based data selection. Also, the proposed algorithm is compared with BSC (a baseline OpenCV all-pairs calibration algorithm) and InterFil (an intermediate filtered variant) using 49 stereo pairs (at 1280 × 720 resolution) captured using a planar checkerboard. OutAw algorithm achieved (using only nine image pairs) superior results (epipolar error 0.5119 px, stereo RMS 0.7666 px) to the BSC ones (epipolar error 1.3687 px, stereo RMS 1.9385 px), representing statistically significant improvements (60.5%, respectively 62.3%). OutAw geometric consistency was validated by triangulation-based metrics (square-length standard deviation 0.1140 mm and square absolute error 0.1097 mm). Contamination analysis revealed that as the outlier rate increases, the calibration process degrades progressively. Also, the results obtained highlight that geometric quality-driven image selection is critical for achieving a reliable stereo calibration for DT applications.

Journal of Imaging doi: 10.3390/jimaging12070279

Authors: Mahzaib Khalid Fangli Ying Al-Garadi Ahmed Mohammed Atef Aniwat Phaphuangwittayakul Riyad Dhuny

Diffusion-based generative models achieve high-fidelity image synthesis; however, controlling internal representations for abstract visual concepts remains challenging due to the ambiguity of textual descriptions. In this work, we propose a prompt-guided concept-vector learning framework for the controllable manipulation of such concepts without requiring external human-annotated image pairs, segmentation masks, identity labels, or manually annotated editing targets. The method introduces a learnable concept vector optimized in the bottleneck (mid-block) feature space of a pretrained Stable Diffusion U-Net, while keeping all pretrained model parameters frozen. A multi-prompt data generation strategy based on paired positive and neutral prompts provides weak semantic guidance for capturing the target concept direction and reducing dependence on a single prompt formulation. The learned vector is further applied in an image-to-image setting through controlled noise injection and concept-guided denoising, enabling the semantic modification of real images while preserving structural content. The concept strength is controlled by a scaling parameter γ, while the image-to-image noise strength is controlled by β, allowing for a practical balance between semantic modification and structural fidelity. Experiments are conducted on two main abstract concepts, perfect skin and peaceful lake, with additional qualitative analysis on subjective portrait-level concepts. Quantitative evaluation using SSIM, LPIPS, and CLIP similarity demonstrates that the proposed method improves semantic alignment while maintaining structural preservation compared with Stable Diffusion image-to-image baselines. A human preference study further shows that concept-injected outputs are preferred in 76.0% of responses for perfect skin and 85.7% for peaceful lake. Ablation studies further demonstrate the controllability and robustness of the proposed framework. Overall, the method provides a simple and parameter-efficient approach for interpretable concept-level manipulation in diffusion models.

Journal of Imaging doi: 10.3390/jimaging12070278

Authors: Ruishi Lin Liyong Ma

Accurate polyp segmentation in colonoscopic images remains challenging due to low contrast, irregular morphology, and significant distribution shifts across datasets, which often lead to unreliable boundary delineation and poor generalization. Existing methods typically treat boundary information as an auxiliary cue or incorporate boundary information through hand-crafted architectural designs, resulting in limited integration between boundary-sensitive features and region-aware representations. In this paper, we propose a boundary-aware multi-task learning framework, termed MBRSNet, which explicitly models and exploits the complementarity between the segmentation task and the auxiliary signed distance field (SDF) regression task. Specifically, we formulate boundary modeling as an auxiliary SDF regression task, providing dense and continuous structural supervision without requiring additional annotations. To effectively couple the two tasks, we design a cross-gated multi-task bottleneck that enables bidirectional and selective feature interaction, allowing each task to selectively leverage complementary information while suppressing task-irrelevant responses. Furthermore, a hierarchical cross-task guidance strategy is introduced in the decoding stage, where boundary-aware weighting and segmentation-guided alignment jointly refine multi-scale features, ensuring consistent integration of boundary cues and regional semantics. Extensive experiments on five benchmark datasets demonstrate that MBRSNet achieves competitive or superior performance compared with representative state-of-the-art methods in both segmentation accuracy and cross-dataset generalization. In particular, the proposed framework achieves superior boundary delineation under challenging conditions and exhibits strong robustness to domain shifts, highlighting the effectiveness of structured task interaction for boundary-aware medical image segmentation.

Journal of Imaging doi: 10.3390/jimaging12070277

Authors: Mahmoud Tahmasebi Saif Huq Kevin Meehan Marion McAfee

Stereo matching has become an increasingly important component of modern autonomous systems. Developing deep learning-based stereo-matching models that deliver high accuracy while operating in real time continues to be a major challenge in computer vision. In the domain of cost volume-based stereo matching, accurate disparity estimation depends heavily on large-scale cost volumes. However, such large volumes store substantial redundant information and also require computationally intensive aggregation units for processing and regression, making real-time performance unattainable. Conversely, small-scale cost volumes followed by lightweight aggregation units provide a promising route for real-time performance, but lack sufficient information to ensure highly accurate disparity estimation. To address this challenge, we propose the Enhanced Shuffle Mixer (ESM) to mitigate information loss associated with small-scale cost volumes. ESM restores critical details by integrating primary features into the disparity upsampling unit. It quickly extracts features from the initial disparity estimation and fuses them with image features. These features are mixed by shuffling and layer splitting, then refined through a compact feature-guided hourglass network to recover more detailed scene geometry. The ESM focuses on local contextual connectivity with a large receptive field and low computational cost, leading to improved disparity estimation accuracy while maintaining real-time performance under the evaluated settings. The compact version of ESMStereo achieves an inference speed of 116 FPS on RTX 4070S and 91 FPS on the AGX Orin.

Journal of Imaging doi: 10.3390/jimaging12070276

Authors: Shiyi Xu Mohan Xu Erjin Zhou

Transformer-based deformable brain MRI registration achieves high overlap accuracy, but predicted displacement fields can contain voxels with a non-positive Jacobian determinant—local foldings that violate the diffeomorphism assumption required by tensor-based morphometry and atlas-fusion segmentation workflows. We introduce HypEReg, a non-linear hyperelastic regularizer that acts directly on the Jacobian determinant of the predicted displacement field. HypEReg couples a clamped-rational volume-distortion penalty (detJϕ−1)2/max(detJϕ,ϵ) with an explicit per-voxel anti-folding hinge [max(0,ϵ−detJϕ)]2, integrated as a purely loss-side module into a TransMorph backbone with no inference-graph modifications. On the IXI atlas-to-subject benchmark (115 test subjects), HypEReg-TransMorph maintains grouped Dice (0.7537) while reducing the det(Jϕ)≤0 voxel ratio from 1.502×10−2 (TransMorph) to 1.5×10−5, with identical per-case runtime and parameter count to the unregularized baseline. In strict zero-shot transfer to OASIS Learn2Reg test pairs (no fine-tuning), HypEReg-TransMorph achieves Dice 0.7756 with a det(Jϕ)≤0 ratio of 7.6×10−5, roughly two orders of magnitude below plain TransMorph zero-shot (Dice 0.7691; ratio 9.6×10−3); downstream multi-atlas label fusion further confirms the practical benefit of fold suppression (fused Dice 0.8271 vs. 0.8201 for TransMorph). OASIS-2 longitudinal and ROI analyses support deformation plausibility (lower folding/SDlogJ and stronger ventricular ROI agreement), while clinical-covariate associations remain exploratory rather than biomarker-validating. Determinant-level, non-linear hyperelastic regularization substantially suppresses folding in Transformer dense-flow brain MRI registration while preserving alignment accuracy and adding zero inference cost, providing a practical drop-in regularization strategy that improves the reliability of deformation fields for morphometry-oriented deformable registration.

Journal of Imaging doi: 10.3390/jimaging12070275

Authors: Maqsudur Rahman Rui Liu Anup Majumder Pintu Chandra Paul Kangtong Mo Amena Begum Kashmi Sultana Nahida Akter Lu Wei Ye Zhang Jun Zhuang

Barren plateaus (BPs) pose a major trainability challenge for quantum neural networks (QNNs) by causing gradients to concentrate near zero as circuit size, depth, or expressibility increases. This study presents a comparative benchmark of 10 BP mitigation strategies across six qubit settings (2, 4, 8, 12, 16, and 20) and three datasets of increasing complexity: Iris, MNIST, and MedMNIST. The evaluated methods include eight initialization-based strategies (Beta, Gaussian, Uniform Norm, CNN-based initialization, He-normal, He-uniform, Xavier-normal, and Xavier-uniform), one model-based variational encoder, and one optimization-based time-nonlocal Fourier parameterization. Experiments were implemented using PennyLane 3.10 and PyTorch 2.5 with simulator backends. We evaluate trainability using gradient variance and training loss, and we clarify that the benchmark analyzes simulated QNN optimization behavior rather than hardware-noise-resilient or noisy-label learning. Across the tested two-layer circuit configurations, the mitigation strategies maintained measurable gradient variance and stable loss reduction, suggesting that severe barren plateau behavior was not observed under the benchmark conditions. CNN-based and Beta initialization showed strong empirical behavior in variance retention and convergence speed, while Gaussian initialization was comparatively weaker in higher-dimensional settings. The study provides a reproducible benchmark structure for comparing BP mitigation behavior and identifies important limitations related to circuit depth, hardware noise, feature encoding, and classification performance that should be addressed in future QNN benchmarking.