Search Results (46)

The transition toward urban-scale digital twin and smart city management requires survey-grade 3D basemaps, yet conventional documentation remains time-consuming and prone to inaccuracies. While Terrestrial Laser Scanning (TLS) offers rapid 3D acquisition, capturing complex, GNSS-denied multi-story interiors frequently causes cumulative registration errors and isolated indoor–outdoor data silos. To address this, this study proposes a comprehensive typology-agnostic framework for developing high-fidelity digital twin basemaps. Treating the building as a unified spatial network, the methodology systematically mitigates error propagation through strategic linkage planning, rigid shell-first registration, continuous vertical core anchoring (via stairwells), and adaptive multi-source data fusion. Implemented on an adaptively reused heritage building, the developed basemap achieved an absolute georeferencing accuracy of 30.0 mm (RMSE) against an independent total station control network, alongside a mean relative error of 3.11 mm. Comparative analysis against legacy 2D CAD floor plans revealed simplified geometric representations and categorical dimensional deviations of up to 29.8 cm. Demonstrating its practical utility, the point cloud-centric geometric hub avoids forced geometric idealization, successfully supporting direct immersive visualization, architectural slicing (floor plans, sections, elevations), and multi-LOD algorithmic planar segmentation. This spatially constrained acquisition strategy bypasses legacy limitations, delivering a mathematically verified 3D reality capture essential for smart facility management, heritage conservation, and downstream semantic intelligence. Full article

Background/Objectives: Accurate delineation of brain tumours on contrast-enhanced MRI remains difficult because lesions can be small, irregular, and weakly separated from adjacent tissue. This study developed BG-YOLO11s, a boundary-guided single-stage instance-segmentation model for T1 contrast-enhanced MRI. Methods: The public Figshare/Cheng dataset, comprising 3064 slices from 233 patients, was converted to YOLO polygon annotations and evaluated using a fixed 70/15/15 image-level split (2144/459/461 slices). Because patient identifiers were not retained in the exported image-and-polygon data, the split was not guaranteed to be patient-disjoint. Bézier Contour Augmentation generated two contour-perturbed training samples per original slice while leaving validation and test data unchanged. BG-YOLO11s extended YOLO11s-seg with dilated context aggregation in the backbone, boundary-enhanced feature fusion in the neck, and a prototype refinement module with differentiable boundary-aware supervision in the segmentation head. Results: In a single run on the held-out image-level test split, BG-YOLO11s achieved 92.4% precision, 88.7% recall, 94.6% mask mAP@50, 68.9% mAP@50–95, and 86.5% IoU. Relative to YOLO11s-seg, the corresponding gains were 3.8 points in mAP@50, 5.8 points in mAP@50–95, and 4.1 points in IoU. A progressive ablation produced incremental gains along the fixed module-addition sequence, but it did not isolate all component interactions or quantify run-to-run uncertainty. Conclusions: BG-YOLO11s improved single-run mask-overlap estimates under the present image-level benchmark. Patient-disjoint retraining, repeated-seed statistics, boundary-specific metrics, complete failure pattern auditing, and external multi-sequence validation are required before broader clinical or deployment claims can be made. Full article

Underwater laser range-gated imaging (ULRGI) effectively suppresses water backscattering via time-slicing mechanisms, making it a primary modality for underwater vision. However, factors such as the inherent optical properties of water, intra-slice residual scattering, gating timing errors, and sensor noise make it difficult to separate target signals from the background. Consequently, the resulting images are generally affected by texture degradation and low contrast, severely limiting the accuracy of downstream tasks like object detection and environmental perception. To this end, we propose the use of a Wavelet-guided Multi-stage Gated Enhancement Network (WMGNet). Operating progressively across three stages, WMGNet’s first two stages employ an encoder–decoder architecture that leverages multi-scale frequency decomposition in the wavelet domain to pinpoint intra-slice scattering and decouple target signals from noise. To precisely extract fine details, we design a TextureBlock integrating feature gating (ConvGLU) and high-frequency attention (HFAttention). Additionally, a pixel-wise ground-truth guided attention module (GGAM) is introduced to optimize the precision and target-specificity of multi-stage feature fusion. Extensive comparative and ablation experiments demonstrate that the proposed WMGNet effectively eliminates scattering interference and restores texture details in underwater imaging. On our custom ULRGI dataset, it achieves state-of-the-art performance with a PSNR of 36.31 dB, an SSIM of 0.921, an MAE of 2.672, and an LPIPS of 0.060. Notably, it outperforms the second-best method by a margin of 3.06 dB in PSNR and reduces the MAE by 50.69%. Furthermore, evaluations on three public datasets confirm its robust cross-scenario generalization, yielding competitive PSNR values of 33.22 dB, 31.59 dB, and 32.06 dB, respectively. Overall, WMGNet provides a highly effective and robust solution for high-resolution underwater imaging. Full article

Porosity-related quantities from industrial X-ray CT depend on segmentation and inference choices. When inference defaults are omitted from the report, void or phase fractions can shift by amounts comparable to slice-to-slice variability. The contribution is metrological rather than architectural: we document a reproducible nnU-Net 2D workflow on Dataset601 CTVoid from semantic labels to slice-wise void fraction, optional two-dimensional connected-component pore summaries, isotropic three-dimensional stacking at 0.058 mm spacing, and spatial axis diagnostics, with region of interest and voxel spacing stated explicitly. The main results pair a weak export policy, defined as a single forward pass per slice without multi-scale fusion or test-time augmentation, with a strong policy that enables multi-scale fusion and flip-based augmentation on the same slice exports and identical weights, on one hundred consecutive slices from one cementitious industrial stack of 1028 × 1028 pixels. In parallel we report trainer validation on eight named Dataset601 validation cases and mirroring-based test-time augmentation off versus on re-inference on those same cases; case identifiers and the cross-validation split appear in the main text. These quantities answer different questions and must not be substituted for one another or for independent full-stack ground truth. Porosity-related scalars from industrial X-ray CT depend on how segmentation and inference are configured; when defaults are omitted, void fractions can shift by amounts comparable to slice-to-slice variability. For fixed nnU-Net weights on one cementitious industrial slice stack (1028 × 1028 pixels), we benchmark weak inference (single forward pass, no multi-scale fusion or test-time augmentation) against a strong export policy (multi-scale fusion and flip-based augmentation) on 100 paired slices, and report parallel trainer validation and TTA-off versus TTA-on re-inference on eight Dataset601 hold-out cases. For the industrial dataset, mean void-class IoU between modes is 0.716 (SD 0.043), while strong inference is ~2.6× slower and predicts lower mean void area (2.37% vs. 3.04%). The full weak export gives a 3D void ratio of 2.44% and integrated void volume of 5175 mm³. On validation patches, mean void Dice/IoU against the reference are 0.835/0.728, while weak–strong void IoU reaches 0.924 under the nnU-Net-native TTA contrast—quantities that must not be interchanged across domains or definitions. The present benchmark does not include a systematic polymer dosage series, and the study does not equate semantic void with open porosity but provides a reproducible disclosure template relevant to porous and polymer-modified cementitious CT reporting. Full article

Pathological complete response (pCR) after neoadjuvant chemotherapy (NACT) provides an endpoint for treatment evaluation in breast cancer. Multi-sequence breast MRI can support pCR prediction, but routine examinations may lack usable T1-weighted or T2-weighted sequences. Many models merge radiomic and deep features by concatenation, leaving the interaction between handcrafted descriptors and learned representations weakly specified. We developed a radiomics-guided framework for pCR prediction from multi-sequence breast MRI. The model uses a multi-branch 2.5D encoder for sequence-specific features, radiomics-guided channel recalibration, and masked token fusion to aggregate available sequence tokens. We evaluated the framework on 157 patients from the I-SPY1 Trial cohort with patient-level five-fold cross-validation, fixed sequence-combination analysis, and slice-window sensitivity analysis. The full model achieved 78.4% accuracy and 0.809 AUC, compared with 75.8% accuracy and 0.788 AUC for the strongest channel-concatenation baseline. In this cohort, radiomics-guided multi-sequence learning was feasible, with external validation required before clinical interpretation. Full article

Background/Objectives: Brain magnetic resonance imaging (MRI) is an important imaging modality for assessing neurological disorders. However, automatic multi-class MRI classification remains challenging because of visual similarity between disease categories, heterogeneous pathological patterns, class imbalance, and the need for reliable confidence estimation. This study aims to develop a comprehensive and well-calibrated deep learning framework for image-level brain MRI classification across multiple neurological categories. Methods: This paper introduces a new deep learning framework, MCND-ComputeNet++, for brain MRI classification into eight image-level categories using the MCND dataset, which comprises 16,400 two-dimensional brain MRI images belonging to eight diagnostic categories: AD-MildDemented, AD-ModerateDemented, AD-VeryMildDemented, BT-glioma, BT-meningioma, BT-pituitary, MS, and Normal. The proposed model uses a single pretrained EfficientNetV2-S backbone to extract hierarchical feature maps from three intermediate stages. These multi-level features are projected into a common latent space, spatially aligned, adaptively fused through learnable gated multi-scale fusion, further refined using convolutional processing, and aggregated using spatial attention pooling before classification. The training strategy combines class-balanced focal loss with label smoothing, MixUp/CutMix regularization, exponential moving average weight smoothing, warmup cosine learning-rate scheduling, temperature scaling, and test-time augmentation to improve generalization and calibration. The framework was evaluated using accuracy, precision, recall, macro-F1, macro-AUC, macro-average precision, expected calibration error, Brier score, bootstrap confidence intervals, ablation analysis, McNemar testing, and comparisons against standard pretrained baseline models. Results: MCND-ComputeNet++ achieved mean accuracy, macro-F1, macro-AUC, and macro-average precision values of 0.9738, 0.9771, 0.9993, and 0.9971, respectively, with narrow bootstrap confidence intervals indicating stable image-level performance. These findings should be interpreted as image-level/slice-level performance on MCND, because patient-level identifiers and subject-wise splitting were not available. These results outperformed most evaluated baselines, including ResNet50, DenseNet121, EfficientNetB0, EfficientNetV2-S with a standard classifier, Swin-Tiny, and ConvNeXt-Tiny, across several discrimination and calibration metrics. Compared with ConvNeXt-Tiny, the proposed model achieved higher macro-AUC and macro-average precision, together with a lower ECE and Brier score, suggesting improved image-level discrimination and confidence reliability. Compared with the EfficientNetV2-S standard classifier, accuracy increased from 0.9308 to 0.9738, while the Brier score decreased from 0.1045 to 0.0400. Conclusions: The results suggest that MCND-ComputeNet++ is a promising image-level brain MRI classification framework for the eight MCND categories. The proposed model integrates hierarchical feature extraction, shared latent projection, gated multi-scale fusion, convolutional refinement, spatial attention pooling, and calibrated inference within a unified architecture. However, because the current evaluation was conducted at the image/slice level without available patient-level identifiers, the findings should not be interpreted as patient-level clinical diagnostic validation. Further studies using subject-wise splitting, external multi-center datasets, 3D volumetric modeling, and multimodal clinical information are required to assess generalizability and potential clinical decision-support applicability. Full article

Alzheimer’s disease (AD) is a global health challenge requiring early and accurate diagnosis, yet current clinical methods struggle with early stages. Deep learning approaches for MRI-based diagnosis face persistent challenges related to image quality issues, limited model generalization, and subtle inter-class variations. To address these limitations, this paper proposes a robust, end-to-end brain MRI-based framework for multi-class classification of AD stages. Positioned within the broader research priority of artificial intelligence and intelligent healthcare technologies, the proposed methodology incorporates an attention-based ensemble of deep learning models alongside an enhanced image preprocessing that uses Real-ESRGAN to mitigate common compression and resolution degradations in 2-D MRI slices. The ensemble makes use of the superior capabilities of the Swin Transformer to capture global contextual dependencies and EfficientNet-B3/MobileNetV2 for effective multi-scale feature extraction, with feature fusion performed using a Squeeze-and-Excitation attention mechanism. The experiments were performed on a publicly available Alzheimer’s MRI dataset, resulting in classification accuracy of 94.47% and 92.28% for the two proposed frameworks. The robustness and clinical interpretability of the framework are emphasized through comprehensive metrics and qualitative analysis. This framework demonstrates promising benchmark performance on a standardized public dataset, highlighting the potential of cross-paradigm ensembles combined with super-resolution preprocessing. Full article

In recent years, the YOLO series has emerged as a widely adopted framework for real-time object detection because of its favorable balance between detection accuracy and inference efficiency. Nevertheless, accurate recognition and localization of tiny defects in industrial inspection remain challenging. These challenges mainly arise from the extremely small scale of defect targets, low image contrast, and the limited capability of conventional models in feature representation under uniform backgrounds. To address these issues from a mathematically optimized perspective and via feature modeling optimization, we develop a dedicated framework for tiny defect detection, termed SLA-YOLO. The main contributions of this work are as follows. First, we adopt a slicing-based processing strategy inspired by the SAHI framework, referred to as Image Slicing Processing (ISP) in this work, and extend it to both training and inference stages. This design enhances the relative scale of tiny defects within local regions, improving detection sensitivity and data diversity without introducing additional model complexity. Second, we introduce a Large Receptive-Field Selective Context (LRSC) module. By leveraging large-receptive-field selective convolution kernels, this module adaptively captures contextual information around critical defect regions via feature modeling optimization of scale-dependent representations. Third, we incorporate a Transformer-based High-level Feature Enhancement (THFE) module to improve global dependency modeling in high-level semantic representations, thereby enhancing feature discriminability for complex defect patterns. Experimental results on the CCB defect dataset show that SLA-YOLO improves mAP@50:95 by 2.7% and mAP@50 by 3.3%. In addition, the proposed method demonstrates strong generalization capability on other tiny object detection tasks. Full article

Optical and Synthetic Aperture Radar (SAR) image registration is a fundamental task in remote sensing information fusion, yet it remains challenging due to significant differences in imaging mechanisms, radiation characteristics, and noise properties between the two modalities. Existing public datasets suffer from limited resolution, small scale, and insufficient scene diversity, and these limitations have hindered algorithm development. This paper constructs a large-scale, high-resolution optical–SAR registration dataset based on the HongTu-1 satellite 3-m SAR imagery and Google Earth optical imagery at zoom level 17, covering diverse scenes across China with a standardized pipeline including terrain correction, geometric alignment, standardized slicing, and quality filtering. Building upon this dataset, a hand-crafted keypoint-based cross-modal registration method is proposed, incorporating multi-level edge filtering and hybrid feature detection. Unlike conventional hand-crafted methods such as RIFT, SRIF, and LNIFT, which mainly refine keypoint detection, description, or matching within a SIFT-style pipeline, the core novelty of this work lies in SAR-specific preprocessing and multi-level hybrid filtering. These components are designed to suppress speckle while extracting more stable and discriminative shared edge responses for cross-modal registration. An improved Log-domain Total Variation (Log-TV) denoising model is introduced for SAR preprocessing. A hybrid edge filtering framework combining phase congruency analysis and Structured Random Forest (SRF) edge detection is constructed within a Gaussian scale space. A dual-branch feature detection scheme integrating blob and corner features is designed with a robust orientation assignment strategy. Feature description uses the Gradient Location–Orientation Histogram (GLOH) descriptor with Principal Component Analysis (PCA) reduction, while geometric estimation employs the Fast Sample Consensus (FSC) algorithm. Experiments on the self-constructed HT dataset and on the public OSdataset and SAR2Opt benchmarks show that the proposed method consistently achieves low RMSE and high success rates. It also maintains competitive efficiency among hand-crafted methods while retaining strong robustness to scale and rotation variations. Full article

Background/Objectives: Breast cancer remains a leading cause of cancer-related mortality among women worldwide. This study presents a systematically justified multi-modal breast cancer classification pipeline that combines established, physically motivated preprocessing operations, modality-specific deep learning models, late-fusion inference, and a deployment-aware federated learning evaluation. Rather than introducing new image restoration or federated optimization algorithms, this work formalizes how standard preprocessing methods can be organized according to the dominant degradation characteristics of ultrasound, MRI, and mammography, and evaluates their contribution under centralized and simulated federated learning settings. Methods: Patient-wise stratified five-fold cross-validation was applied across ultrasound (BUSI, $n=780$), dynamic contrast-enhanced MRI (DUKE, $n=922$), and mammography (CBIS-DDSM, $n=400$). A five-algorithm federated learning comparison, including FedAvg, FedProx, SCAFFOLD, FedNova, and FP16-FedAvg, was conducted under IID and non-IID conditions using a Dirichlet distribution with $\alpha =0.5$. The evaluation reports diagnostic performance together with per-round training time, communication time, latency-related measurements, and cumulative bandwidth. Ablation experiments, McNemar’s test, Cohen’s h effect sizes, and confidence intervals were used to support the analysis. Results: Per-modality models achieved 92.50 ± 1.2%, 90.63 ± 1.5%, and 92.00 ± 1.3% accuracy for ultrasound, MRI, and mammography, respectively, with statistically significant improvements over the corresponding baselines according to McNemar’s test ($p<0.05$). Weighted late fusion achieved 93.10 ± 1.1% accuracy and improved performance compared with the best individual modality ($p=0.031$). FP16 transmission reduced cumulative bandwidth from 8.14 GB to 1.23 GB ($-84.9\%$) without a statistically significant performance difference compared with FP32 transmission ($p=0.74$), while SCAFFOLD achieved the highest non-IID accuracy (90.50%). Conclusions: The findings demonstrate internal technical validity and deployment-relevant trade-offs, but they should be interpreted cautiously because the federated evaluation is simulation-based, key-slice extraction may require annotation-assisted assumptions, and external multi-center validation remains necessary before clinical deployment. Reported improvements are statistically significant in several comparisons, but corresponding Cohen’s h effect sizes are small, and clinical meaningfulness requires independent validation rather than inference from p-values alone. Full article

Accurate multi-class segmentation of the prostate in T2-weighted magnetic resonance imaging (MRI) into the peripheral zone (PZ), central gland (CG) and tumour is essential for targeted biopsy guidance and treatment planning. We present DSBANet, an encoder–decoder architecture that combines a pretrained ResNet-50 encoder, Atrous Spatial Pyramid Pooling, Multi-Scale Attention Fusion on skip connections, a Feature Fusion Module, deep supervision and boundary refinement. We evaluate eight architectures across three input dimensionalities (2D, 2.5D, 3D), yielding 24 models trained under identical conditions on the Prostate158 dataset. DSBANet achieves the best anatomy segmentation with PZ DSC of 0.8176 and CG DSC of 0.7888 among 2D models. To address the severe class imbalance of the tumour class, we further train DSBANet 2D with a class-weighted cross-entropy term and tumour-positive slice oversampling, raising per-case tumour DSC from 0.003 to 0.170 (a sixty-fold absolute improvement). A systematic eight-variant ablation study, evaluated under matched-pairs effect-size analysis, identifies the SE-Residual blocks and skip-connection attention as the largest contributors to tumour segmentation, while every architectural component contributes a directionally consistent gain. Full article

Objectives: To exploratorily evaluate the potential of baseline dedicated breast PET (D-PET) for noninvasive prediction of pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in HER2-positive (HER2+) breast cancer, and to investigate a fusion strategy integrating conventional radiomics and deep learning features. Methods: We developed a multi-representation framework with radiomics based on data-driven high-/low-uptake metabolic subregions and deep learning trained on standardized 3D tumor volumes, and intratumoral heterogeneity (ITH) was quantified on the largest slice as an additional comparator. The outputs of these pathways were subsequently integrated through feature-level and decision-level fusion. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), and interpretability analyses were applied to identify image regions and features contributing to predictions. Results: In a HER2-positive breast cancer cohort (n = 147) with baseline D-PET, deep learning (3D ResNet, AUC = 0.79) and radiomics (logistic regression, AUC = 0.78) achieved comparable performance on the primary test set, whereas the ITH model showed limited value (AUC = 0.61). Fusion further improved discrimination on test set 1, with an AUC of 0.83 for decision-level fusion and 0.84 for feature-level fusion. On test set 2, decision-level fusion achieved the highest AUC (0.84), and feature-level fusion maintained stable performance (AUC = 0.80). Conclusions: In this exploratory study, baseline D-PET showed promising performance for noninvasive prediction of NAC response in HER2+ breast cancer. The fusion of deep learning and radiomics yielded improvements over single-representation models, highlighting the potential role of D-PET models as decision-support tools. Full article

Shape from Focus (SFF) estimates scene depth by analyzing focus variations across a sequence of images captured at different focal settings. Traditional SFF methods rely on handcrafted focus operators that preserve local structural details, but they are often sensitive to noise and perform poorly in textureless regions. In contrast, deep learning-based methods are more robust and can exploit semantic and contextual cues, yet they may lose fine structural information due to feature abstraction and spatial downsampling. To address these complementary limitations, we propose a dual-branch SFF framework that integrates deep and traditional focus cues within a unified architecture. The first branch generates a deep focus volume using a multi-scale encoder-decoder network, while the second branch computes a traditional focus volume using a directional dilated Laplacian (DDL) operator to capture structural focus responses. These two volumes are progressively combined through an iterative gated fusion module, producing a more discriminative fused focus representation. From this fused volume, an initial depth map is estimated through a softmax-based slice aggregation strategy. To further improve spatial consistency and reduce residual artifacts, we introduce a lightweight depth refinement module guided by the mean RGB image of the focal stack. This refinement stage enhances boundary quality and improves the overall depth structure. Extensive experiments on synthetic and real-world datasets demonstrate that the proposed framework produces accurate and reliable depth maps. Full article

Magnetic resonance imaging (MRI) remains one of the most important tests for diagnosing and monitoring various diseases. In recent years, machine learning methods have been widely applied to automate MRI analysis. It supports decision-making by predicting disease and highlighting relevant regions. However, the proper use of feature extraction methods can improve the performance of the model. This paper proposes a WaveletFusion architecture that combines a two-dimensional Haar wavelet decomposition with a convolutional neural network (CNN) for classification. The approach was demonstrated on the Brain Tumor MRI dataset and further examined on the Br35H :: Brain Tumor Detection 2020 (Br35H). The model decomposes each MRI slice into approximation and directional detail subbands and fuses multi-scale wavelet features within the convolutional pipeline. To evaluate the effect of decomposition depth, WaveletFusion variants from one to eight levels were compared with a Baseline CNN model under the same training protocol. The results showed that performance improved progressively with increasing decomposition depth up to level 7, whereas the 8-level configuration consistently declined, indicating that excessive decomposition introduces information loss and over-compression in the deepest approximation pathway. The best-performing configuration, which outperformed both the Baseline CNN and the WaveletFusion variations in five independent runs, was the 7-level WaveletFusion model, achieving a test accuracy of 0.94 ± 0.01 and test macro-F1 of 0.93 ± 0.02. A similar tendency was observed on the Br35H dataset, where the 7-level model achieved a 0.97 ± 0.01 test accuracy and 0.97 ± 0.01 test macro-F1, while the 8-level configuration remained weaker on both datasets. These results show that multi-scale wavelet fusion can improve Brain Tumor MRI classification while maintaining a compact model size and a fair comparison setting, and that the decomposition depth must be selected carefully. Full article

Instance segmentation of plant point clouds is challenging due to intricate structures, non-uniform density, and large intra-class variation. Conventional methods often suffer from blurred boundaries, instance adhesion, and insufficient coupling of semantic and instance features. To address these issues, this paper proposes MPRSF-CSA, a novel network integrating recursive slice-based feature extraction with an attention-embedding mechanism. The method first transforms disordered point clouds into ordered sequences via a multi-directional recursive slicing strategy and models inter-slice dependencies using BiLSTM. Parallel decoding branches for semantic and instance segmentation are constructed, and a core attention-embedding module facilitates bidirectional fusion of semantic and instance features. Instance segmentation is achieved via clustering and semantic-aware optimization. Experiments on two public datasets demonstrate that MPRSF-CSA outperforms existing approaches in segmentation accuracy, boundary preservation, and adaptability to complex plant scenes. Full article