Diagnostics

Research

Jump to: Review, Other

13 pages, 1923 KB

Open AccessArticle

AI–Enabled Interpretation Guidance for Hemostasis Testing with TEG^® 6s

by Jan Hartmann, Dana Souter, Joao D. Dias and Qun Sha

Diagnostics 2026, 16(17), 2852; https://doi.org/10.3390/diagnostics16172852 - 4 Sep 2026

Viewed by 326

Abstract

Background/Objectives: The TEG^® 6s is a hemostasis analyzer system used to assess viscoelastic properties of whole blood. Although it provides rapid and comprehensive insights, interpreting thromboelastography tracings requires robust clinical training and is not standardized. Our aim was to develop and [...] Read more.

Background/Objectives: The TEG^® 6s is a hemostasis analyzer system used to assess viscoelastic properties of whole blood. Although it provides rapid and comprehensive insights, interpreting thromboelastography tracings requires robust clinical training and is not standardized. Our aim was to develop and evaluate a prototype of MetaTEG—an AI-driven support tool for interpreting TEG^® 6s tracings, with a focus on the Global Hemostasis-Heparin Neutralization cartridge. Methods: TEG^® 6s tracings, obtained from Haemonetics Corporation’s internal case library, were used to train a custom AI agent developed with Microsoft Copilot Studio. Five different clinical cases were used for testing. The AI agent incorporated knowledge graph augmentation and prompt engineering to integrate published literature and assess coagulation states based on R-time, maximum amplitude, and other key parameters. The evaluation of the AI virtualization outcomes included semiquantitative scoring by an internal TEG expert panel who assessed the AI’s diagnostic accuracy, tracing interpretation, and therapeutic recommendations. UI/UX prototypes were designed in Figma to demonstrate potential integration into clinical workflows. Results: Using this proof-of-concept dataset and qualitative expert scoring, MetaTEG accurately described and interpreted TEG^® 6s tracings and suggested potential treatment options. Reviewers consistently rated the AI-generated outputs as both clinically useful and accurate. Conclusions: The MetaTEG prototype represents the first application of generative AI for TEG^® 6s interpretation, showcasing the potential of AI-powered, human-in-the-loop decision support in hemostasis diagnostics. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

34 pages, 4725 KB

Open AccessArticle

A Resource-Efficient Hybrid Deep Learning Framework with External Validation for Intracranial Hemorrhage Detection in CT Scans

by José Rafael Peña Gutiérrez and César Julio Bustacara Medina

Diagnostics 2026, 16(17), 2776; https://doi.org/10.3390/diagnostics16172776 - 29 Aug 2026

Viewed by 389

Abstract

Background: Intracranial hemorrhage (ICH) is a time-critical neurological emergency in which delayed diagnosis significantly worsens patient outcomes. This challenge is amplified in resource-limited, high-workload settings where rapid neuroimaging interpretation may be constrained. Many high-performing deep learning approaches for disease detection rely on large-scale [...] Read more.

Background: Intracranial hemorrhage (ICH) is a time-critical neurological emergency in which delayed diagnosis significantly worsens patient outcomes. This challenge is amplified in resource-limited, high-workload settings where rapid neuroimaging interpretation may be constrained. Many high-performing deep learning approaches for disease detection rely on large-scale models trained on extensive data and evaluated only on internal datasets, limiting their generalization across heterogeneous clinical environments. This study aims to develop and evaluate a resource-efficient framework for automated ICH detection from CT scans, with internal and external validation across heterogeneous clinical settings. Methods: A hybrid deep learning framework was developed, combining an EfficientNetV2-S-based feature extractor with a bidirectional GRU model for scan-level prediction. The model was trained on a stratified subset of 6000 CT scans from the RSNA Intracranial Hemorrhage Detection dataset and evaluated using an internal test set and two external validation cohorts (PhysioNet and CQ500). Results: On an internal held-out test set, the model achieved scan-level AUROC and AUPRC of 0.980 and 0.977, respectively, and slice-level AUROC and AUPRC of 0.981 and 0.923. External validation on the PhysioNet and CQ500 datasets yielded scan-level AUROC/AUPRC values of 0.914/0.932 and 0.905/0.909, respectively, demonstrating consistent performance across datasets differing in institution, geography, patient population, and acquisition protocols. Conclusions: Despite its compact architecture and reduced training subset, the proposed framework achieves performance competitive with substantially larger and more computationally demanding models, completing training in under 26 h on single-GPU hardware. These results support the feasibility of reproducible, resource-efficient ICH detection systems for automated triage in emergency radiology workflows across different clinical settings. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Graphical abstract

30 pages, 10174 KB

Open AccessArticle

Hybrid Vision Transformer–CNN Architecture with Optimized Feature Selection for Skin Cancer Classification

by Abrar Almjally, Munazza Aziz, Shaheryar Najam, Alaa Menshawi, Meteb Altaf, Abdullah Fawaz Aljulayfi and Ahmad Jalal

Diagnostics 2026, 16(15), 2351; https://doi.org/10.3390/diagnostics16152351 - 27 Jul 2026

Viewed by 451

Abstract

Background/Objectives: Melanoma is a life-threatening skin cancer characterized by aggressive progression and high metastatic potential, making early diagnosis essential for improving patient survival and treatment outcomes. However, accurate automated skin lesion classification remains challenging due to variations in lesion appearance, illumination, image quality, [...] Read more.

Background/Objectives: Melanoma is a life-threatening skin cancer characterized by aggressive progression and high metastatic potential, making early diagnosis essential for improving patient survival and treatment outcomes. However, accurate automated skin lesion classification remains challenging due to variations in lesion appearance, illumination, image quality, and the presence of artifacts. This study proposes a unified framework for robust multi-class skin cancer classification by integrating preprocessing, lesion segmentation, feature extraction, optimization, and classification within a single end-to-end architecture. Methods: The proposed framework employs an iterative hair artifact removal strategy based on the fusion of Frangi vesselness filtering, Gabor texture filtering, morphological refinement, and Telea inpainting to preserve lesion integrity. A novel dermoscopic lesion segmentation network (CutisNet) is introduced to accurately delineate lesion boundaries. Hybrid representation learning combines deep features extracted using MobileNetV2 with uniquely selected handcrafted descriptors to capture complementary texture, structural, and contextual information. Gray Wolf Optimization is utilized for feature fusion and refinement, while a hybrid GNN–CNN classifier performs robust multi-class skin lesion classification. Results: Extensive experiments conducted on multiple benchmark dermoscopic datasets demonstrate the effectiveness and generalization capability of the proposed framework. The proposed model consistently outperformed existing state-of-the-art methods, achieving a maximum classification accuracy of 96.7% on the PH² dataset while maintaining competitive performance across other benchmark datasets. Conclusions: The proposed framework effectively integrates novel preprocessing, segmentation, hybrid feature representation, feature optimization, and classification strategies to improve the robustness and accuracy of automated skin cancer classification. These results demonstrate its potential to support reliable computer-aided diagnosis and assist clinicians in the early detection of skin cancer. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

33 pages, 6485 KB

Open AccessArticle

ABMA: An Attention-Based Morphology-Aware Framework for Automated 12-Lead ECG Arrhythmia Classification

by Manjur Kolhar and Raisa Nazir Ahmed Kazi

Diagnostics 2026, 16(14), 2274; https://doi.org/10.3390/diagnostics16142274 - 21 Jul 2026

Viewed by 431

Abstract

Background: Cardiovascular diseases (CVDs) are among the leading causes of death globally. In order to treat CVDs successfully in the early stages, it is crucial to diagnose them in time. The ECG is one of the most common and non-invasive methods to detect [...] Read more.

Background: Cardiovascular diseases (CVDs) are among the leading causes of death globally. In order to treat CVDs successfully in the early stages, it is crucial to diagnose them in time. The ECG is one of the most common and non-invasive methods to detect heart rhythms and to diagnose arrhythmias. However, the analysis of ECG recordings manually requires a lot of time and experience because the morphology of ECG signals and the characteristics of their waveforms are very complex and show large overlaps between different types of arrhythmias. So far, various approaches for automated analysis of ECG signals have been developed, mostly based on deep learning (DL). In general, these methods are able to analyze ECG signals automatically and to detect different types of arrhythmias. Most approaches, however, are based on a purely data-driven feature learning and do not pay attention to the morphology-sensitive temporal structure of ECG signals, which is important for a discriminative diagnosis of arrhythmias. Methods: In this paper, we propose an Attention-Based Morphology-Aware (ABMA) framework to leverage multilead ECG signals in conjunction with automatically computed physiological features using a hybrid deep learning architecture. ABMA leverages multi-scale convolutional neural networks to learn local morphology features, and bidirectional long short-term memory (BiLSTM) networks to model temporal rhythms in ECG signals. We designed an ABMA module that incorporates a morphology scoring network (MSN) in order to (1) estimate the morphology-aware importance of different ECG segments and (2) learn the temporal importance of ECG features. The learned attention weights enable learning to focus on key sections of ECG signals without predefined boundaries or manual annotation of fiducial points. To understand the contribution of each individual component of the framework, we performed an extensive ablation study, where we removed the handcrafted feature branch, the ABMA module, the MSN, and the multi-head attention mechanism, one at a time, and compared the results against a fixed set of experimental configurations. Results: To assess the performance of the proposed framework in three-class classification between sinus rhythm (SA), atrial fibrillation (AFIB), and ventricular tachycardia (VT), we employed a stratified 10-fold cross-validation protocol. Our approach achieved a mean accuracy of 95.18 ± 1.18%, followed by a corresponding weighted F1-score of 95.19 ± 1.18% and a macro F1-score of 94.66 ± 1.35%. Notably, the performance of the proposed complete ABMA framework considerably outperformed the baseline CNN–BiLSTM architecture. Furthermore, in the primary evaluation metrics (i.e., accuracy, F1-score), the complete framework showed statistically significant improvements against the baseline through paired two-sided t-tests (p < 0.001). The ablation study indicated that each architectural component contributed positively to the overall classification performance, with the complete ABMA framework outperforming all reduced variants. Conclusions: The framework was evaluated by stratified cross-validation on a publicly available dataset. Our framework outperformed the baseline CNN–BiLSTM model as well as the respective ablation models in terms of classification performance. The findings from the current study are based on a retrospective analysis and therefore future studies using an independent external dataset, from multiple centers, or as part of a prospective clinical study are necessary in order to establish the generalizability and clinical utility of the proposed framework. The ABMA framework is currently viewed as a very promising research framework for intelligent ECG analysis, but it is not yet a clinically validated diagnostic tool. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

20 pages, 9200 KB

Open AccessArticle

A Multilevel Deep Learning Model for Automated Brain Tumor Segmentation Using Magnetic Resonance Images

by Aneesh S. Perumprath, Kulandairaj Martin Sagayam, Syed Immamul Ansarullah, Shafat Khan, Sami Alshmrany, Farhan Amin, Isabel de la Torre Díez, Mirtha Silvana Garat de Marin and Eduardo Silva Alvarado

Diagnostics 2026, 16(14), 2259; https://doi.org/10.3390/diagnostics16142259 - 20 Jul 2026

Viewed by 531

Abstract

Background/Objectives: Brain tumor segmentation from magnetic resonance imaging (MRI) plays an important role in clinical assessment and treatment planning. However, accurate segmentation remains challenging because of the complex anatomical structure of the brain, variations in tumor size and shape, and the imbalance between [...] Read more.

Background/Objectives: Brain tumor segmentation from magnetic resonance imaging (MRI) plays an important role in clinical assessment and treatment planning. However, accurate segmentation remains challenging because of the complex anatomical structure of the brain, variations in tumor size and shape, and the imbalance between tumor and non-tumor regions in MRI datasets. These challenges highlight the need for reliable automated segmentation methods. Methods: This study proposes a multilevel deep learning model for automated brain tumor segmentation using MRI images. The BraTS dataset was used for model development and evaluation. To address class imbalance, a modified Synthetic Minority Oversampling Technique (SMOTE) was incorporated during preprocessing. A Multilevel Architecture-Based Modified U-Net was then employed to learn multiscale spatial features and generate pixel-wise tumor segmentation. The proposed framework was evaluated using the Dice coefficient, Jaccard coefficient, Matthews Correlation Coefficient (MCC), and accuracy. Results: The experimental results demonstrate that the proposed model consistently outperformed the Berkeley Wavelet Transform (BWT)-based method and the conventional U-Net across different tumor grades. Higher Dice, Jaccard, and MCC values indicate improved agreement between the predicted segmentation and the expert-annotated ground truth masks, demonstrating more accurate and consistent tumor delineation. Conclusions: The proposed multilevel deep learning model provides an effective framework for automated brain tumor segmentation from MRI images. By combining imbalance-aware preprocessing with a lightweight Modified U-Net architecture, the proposed method improves segmentation performance while maintaining a relatively simple network design. Future work will focus on validating the proposed framework using external clinical datasets and comparing it with recent state-of-the-art segmentation models. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

17 pages, 884 KB

Open AccessArticle

Diagnostic Accuracy and Time Efficiency of Artificial Intelligence for Intracranial Hemorrhage Detection on Emergency Brain CT

by Ömer Damar, Mahmut Yaman and Mehmet Tahir Gokdemir

Diagnostics 2026, 16(14), 2219; https://doi.org/10.3390/diagnostics16142219 - 16 Jul 2026

Viewed by 468

Abstract

Background: Intracranial hemorrhage (ICH) is a neurological emergency associated with high mortality and morbidity, requiring rapid diagnosis and early intervention. Artificial intelligence (AI)-based imaging systems have recently gained attention as supportive tools for accelerating emergency brain computed tomography (CT) interpretation. This study [...] Read more.

Background: Intracranial hemorrhage (ICH) is a neurological emergency associated with high mortality and morbidity, requiring rapid diagnosis and early intervention. Artificial intelligence (AI)-based imaging systems have recently gained attention as supportive tools for accelerating emergency brain computed tomography (CT) interpretation. This study aimed to evaluate the diagnostic accuracy and time efficiency of an AI system for detecting ICH on emergency brain CT scans. Methods: This retrospective study included 375 patients who underwent non-contrast brain CT in the emergency department. AI-based image analysis results were compared with final radiologist reports accepted as the reference standard. Sensitivity, specificity, predictive values, accuracy, area under the curve (AUC), and Cohen’s kappa were calculated. Interpretation times, subgroup analyses according to presentation type and work shifts, and hemorrhage volume correlations between AI and the ABC/2 method were also evaluated. Results: Intracranial hemorrhage was detected in 235 patients (62.7%). The AI system achieved a sensitivity of 94.0%, specificity of 87.9%, positive predictive value of 92.9%, negative predictive value of 89.8%, and overall accuracy of 91.7%. The AUC was 0.909, with a Cohen’s kappa coefficient of 0.823. Mean interpretation time was significantly shorter for AI compared with radiologists (11.9 ± 3.0 vs. 70.9 ± 25.4 min, p < 0.001). AI-derived hemorrhage volume measurements strongly correlated with the ABC/2 method (rho = 0.887, p < 0.001). Conclusions: The AI system demonstrated high diagnostic accuracy and a significant reduction in interpretation time for detecting intracranial hemorrhage on emergency brain CT, supporting its potential role as a decision-support tool in emergency radiology workflows. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

22 pages, 1547 KB

Open AccessArticle

Bridging Annotation Gaps: Hierarchical Self-Support Learning for Brain Tumor Segmentation

by Saqib Qamar, Mohd Fazil and Zubair Ashraf

Diagnostics 2026, 16(11), 1588; https://doi.org/10.3390/diagnostics16111588 - 22 May 2026

Cited by 1 | Viewed by 489

Abstract

Background: Accurate brain tumor segmentation from Magnetic Resonance Imaging (MRI) depends on the fusion of multiple complementary modalities. However, clinical practice often faces incomplete modality sets due to acquisition failures, patient contraindications, or protocol variations. Current methods either treat each modality feature extractor [...] Read more.

Background: Accurate brain tumor segmentation from Magnetic Resonance Imaging (MRI) depends on the fusion of multiple complementary modalities. However, clinical practice often faces incomplete modality sets due to acquisition failures, patient contraindications, or protocol variations. Current methods either treat each modality feature extractor in isolation or depend on computationally expensive teacher networks for cross-modal knowledge transfer. Objective: This paper presents Hierarchical Adaptive Group Self-Support Learning with Boundary-Aware Calibration (HAGSS), a framework that overcomes three key limitations of existing group self-support methods: static group formation that ignores temporal prediction quality, uniform treatment of boundary and interior voxels, and distribution mismatch across heterogeneous modality logits. Methods: We propose a hierarchical adaptive group formation mechanism that reassigns group leader roles at each epoch based on voxel-level prediction confidence scores instead of fixed sensitivity priors. We also introduce a boundary-aware calibration module that applies spatially varied distillation weights with greater emphasis on tumor boundary regions. In addition, we design a cross-scale consistency regularization term that enforces agreement between multi-resolution predictions to stabilize the self-support target. Results: Experiments on BraTS2020, BraTS2018, and BraTS2021 datasets show that HAGSS achieves consistent improvements over state-of-the-art baselines. The average Dice gains across the whole tumor, tumor core, and enhancing tumor regions reach 1.30% on BraTS2020 and 1.61% on BraTS2021 compared to existing methods. All improvements are statistically significant (

p < 0.05

). Conclusions: HAGSS operates exclusively during training, adds no parameters or inference cost, and can be applied as a plug-in module to any multi-encoder incomplete multi-modal segmentation architecture. Code is publicly available at GitHub. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

25 pages, 6880 KB

Open AccessArticle

SULBA: A Task-Agnostic Data Augmentation Framework for Deep Learning in Medical Image Analysis

by Ayomide Adeyemi Abe and Mpumelelo Nyathi

Diagnostics 2026, 16(10), 1546; https://doi.org/10.3390/diagnostics16101546 - 19 May 2026

Cited by 1 | Viewed by 596

Abstract

Background/Objectives: Data augmentation is a foundational component of modern deep learning for enhancing robustness and generalization. However, medical imaging lacks a universally reliable augmentation strategy, forcing researchers into an inefficient “augmentation lottery” that hinders experimental progress and reproducibility. To address this challenge, [...] Read more.

Background/Objectives: Data augmentation is a foundational component of modern deep learning for enhancing robustness and generalization. However, medical imaging lacks a universally reliable augmentation strategy, forcing researchers into an inefficient “augmentation lottery” that hinders experimental progress and reproducibility. To address this challenge, we introduce Stepwise Upper and Lower Boundaries Augmentation (SULBA), a simple, parameter-free framework designed to eliminate per-task augmentation tuning. Methods: SULBA generates training variations through stepwise cyclic shifts applied along data dimensions, making it inherently applicable to 2D, 3D, and higher-dimensional medical imaging data. To evaluate the efficacy of SULBA as a default DA strategy, we performed benchmarking across 27 publicly available datasets spanning classification and segmentation tasks and 10 convolutional and transformer-based architectures using standard deep learning performance metrics. Results: The results demonstrate that SULBA achieves the highest overall performance and consistently outperforms 16 widely used standard augmentation techniques while delivering robust and reliable improvements without task- or parameter-specific tuning Conclusions: SULBA establishes a principled universal default for data augmentation in medical imaging, with the potential to accelerate the development of generalizable and reproducible medical AI systems. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

21 pages, 1467 KB

Open AccessArticle

Comparative Analysis of Third Molar Segmentation Performance Between Sexes Using Deep Learning Models

by Ayşe Bulut, Melis Büşra Aşkın and Gökalp Çınarer

Diagnostics 2026, 16(7), 977; https://doi.org/10.3390/diagnostics16070977 - 25 Mar 2026

Viewed by 763

Abstract

Background/Objectives: Sex analysis in dental radiographs plays a central role in forensic identification, especially when biological material is compromised or incomplete. While most AI-based studies rely on complete dentition or craniofacial structures, this study investigates whether sex-based information can be extracted solely through [...] Read more.

Background/Objectives: Sex analysis in dental radiographs plays a central role in forensic identification, especially when biological material is compromised or incomplete. While most AI-based studies rely on complete dentition or craniofacial structures, this study investigates whether sex-based information can be extracted solely through segmentation of third molars in panoramic radiographs. Methods: A retrospective dataset containing 2818 third molar annotations from 757 panoramic images with balanced class distribution across training, validation, and testing subsets was constructed. Three sample segmentation-based deep learning models—YOLOv12n, YOLO26n, and RT-DETR v2—were evaluated under the same training conditions using detection-focused metrics including sensitivity, recall, and mAP. Results: YOLOv12n demonstrated the most balanced performance, achieving the highest mAP@0.50 score of 0.810 and mAP@0.50–0.95 score of 0.574; RT-DETR v2 showed higher sensitivity but lower localization accuracy and significantly longer training time. YOLO26n yielded the highest recall rate but showed an increase in false positives. Class-based analyses indicated sex-specific morphological variability in third molar anatomy, showing consistently higher detection performance in female samples. Conclusions: These results demonstrate that isolated third molars encode distinctive sex-related signals and that segmentation-focused frameworks offer an interpretable and clinically relevant alternative to whole-image classification in forensic dentistry. Future studies should incorporate larger multi-population datasets, multi-tooth integration, and explainable AI techniques to further improve robustness and applicability. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

22 pages, 52674 KB

Open AccessArticle

Lightweight Deep Learning for Automated Dental Caries Screening from Pediatric Oral Photographs

by Nourah Alangari and Nouf AlShenaifi

Diagnostics 2026, 16(6), 862; https://doi.org/10.3390/diagnostics16060862 - 13 Mar 2026

Viewed by 1151

Abstract

Background: Early childhood caries (ECC) affects a substantial proportion of young children worldwide, and timely screening is essential for early intervention and referral. While deep learning has shown promise for automated dental diagnostics, many existing approaches rely on computationally heavy models that limit [...] Read more.

Background: Early childhood caries (ECC) affects a substantial proportion of young children worldwide, and timely screening is essential for early intervention and referral. While deep learning has shown promise for automated dental diagnostics, many existing approaches rely on computationally heavy models that limit deployment in community and mobile settings. This study investigates whether compact convolutional neural networks can achieve clinically meaningful performance for screening dental caries from oral photographs. Methods: We curated a dataset of 435 intraoral images from children aged 3–14 years, annotated by licensed dentists, and performed patient-level stratified splitting to prevent data leakage. Three convolutional neural networks (ResNet-18, MobileNetV3-Small, and EfficientNet-B0) were fine-tuned using ImageNet-pretrained weights and comparatively evaluated for the detection of dental caries from oral photographs. Models were trained with class-weighted cross-entropy loss and evaluated on a held-out test set using sensitivity, specificity, balanced accuracy, ROC-AUC, and PR-AUC with bootstrap 95% confidence intervals. Results: ResNet-18 achieved the highest balanced accuracy (0.929), weighted F1-score (0.954), and perfect sensitivity (1.00), while EfficientNet-B0 achieved the strongest threshold-independent discrimination with the highest ROC-AUC (0.978) and PR-AUC (0.990). MobileNetV3-Small maintained competitive performance (ROC-AUC 0.952; PR-AUC 0.976) with substantially lower computational complexity. Conclusions: In addition to performance evaluation, we incorporated an interpretability analysis using Grad-CAM to examine model decision behavior. The resulting attribution maps predominantly highlighted clinically relevant tooth regions associated with caries, providing evidence that the models rely on meaningful dental features rather than background artifacts. These results demonstrate that compact, deployment-friendly architectures can achieve clinically meaningful performance for ECC detection, supporting their suitability for scalable, real-world screening applications. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

27 pages, 4440 KB

Open AccessArticle

Optimization-Driven Hybrid Machine Learning Framework for Brain Tumor Classification in MRI with Metaheuristic Feature Selection

by Yasin Özkan, Yusuf Bahri Özçelik and Aytaç Altan

Diagnostics 2026, 16(5), 819; https://doi.org/10.3390/diagnostics16050819 - 9 Mar 2026

Cited by 25 | Viewed by 1423

Abstract

Background/Objectives: Brain tumors are among the most severe neurological disorders, and their variability in size, morphology, and anatomical location complicates early and accurate diagnosis. Although magnetic resonance imaging (MRI) is the most reliable non-invasive modality for tumor detection, manual interpretation remains time-consuming, subjective, [...] Read more.

Background/Objectives: Brain tumors are among the most severe neurological disorders, and their variability in size, morphology, and anatomical location complicates early and accurate diagnosis. Although magnetic resonance imaging (MRI) is the most reliable non-invasive modality for tumor detection, manual interpretation remains time-consuming, subjective, and susceptible to human error. This study aims to develop an optimization-driven hybrid machine learning framework for accurate and computationally efficient automatic brain tumor classification. Methods: The dataset includes 834 MRI images (583-training, 123-validation, 128-independent test). Because YOLOv11 detects tumor and non-tumor regions separately, the sample size doubled during region-based analysis, and all subsequent stages were conducted at the regions of interest (ROI) level. On the independent test set, YOLOv11 achieved 98.87% mAP@50, 98.54% precision, and 98.21% recall. The proposed framework combines automated tumor localization with image standardization using Gaussian noise reduction and bilinear interpolation. From the processed MR images, 39 entropy-based features were extracted. To enhance diagnostic performance and eliminate redundant information, the superb fairy-wren optimization algorithm (SFOA) was applied for feature selection and compared with particle swarm optimization (PSO), Harris hawk optimization (HHO), and puma optimization (PO). Final classification was primarily performed using k-nearest neighbors (kNN), while support vector machines (SVM) were used for comparative evaluation. Results: SFOA reduced the feature dimensionality from 39 to 5 features while achieving 99.20% classification accuracy on the independent test set. In comparison, PSO selected 10 features, HHO selected 6 features and PO selected 10 features, all achieving 98.45% accuracy. The best performance obtained with SVM was 98.45% accuracy (HHO-SVM), which remained lower than the 99.20% achieved by the proposed SFOA-kNN model. Conclusions: The results indicate that combining entropy-based feature extraction with SFOA-driven feature selection and kNN classification significantly enhances diagnostic accuracy while reducing computational complexity, highlighting the strong potential of the proposed framework for integration into computer-aided diagnosis systems to support clinical decision-making. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

27 pages, 8529 KB

Open AccessArticle

Ensemble Deep Learning-Based High-Precision Framework for Breast Cancer Detection from Histopathological Images

by Faizan Ahmad, Arfan Jaffar, Ghazanfar Latif, Jaafar Alghazo and Sohail Masood Bhatti

Diagnostics 2026, 16(5), 653; https://doi.org/10.3390/diagnostics16050653 - 24 Feb 2026

Cited by 3 | Viewed by 1180

Abstract

Background/Objectives: Analysis of histopathological images is the absolute standard of breast cancer diagnosis. However, modern deep learning- and ViT-based architecture still struggle to capture effective local and global discriminatory patterns that tend to make architecture more complex, increasing the risk of overfitting [...] Read more.

Background/Objectives: Analysis of histopathological images is the absolute standard of breast cancer diagnosis. However, modern deep learning- and ViT-based architecture still struggle to capture effective local and global discriminatory patterns that tend to make architecture more complex, increasing the risk of overfitting and optimization problems. Methods: To address these problems, this paper proposes a four-phase hybrid framework that aims to enhance the feature fusion, improving the model’s strength, robustness, and generalization ability. In Phase 1, the BreakHis dataset was split patient-wise into a 70-15-15 manner to avoid data leakage, while extensive data augmentation, comprehensive normalization, and a five-fold cross-validation protocol were implemented to make the dataset more varied and reliably evaluated without bias. Phase 2 entailed the training of three CNNs (VGG16, ResNet50, and DenseNet121) and four ViTs (DeiT, CaiT, T2T-ViT, and Swin Transformer) independently to establish the strict baseline performance standards. In Phase 3, the CNN-based features were fused and classified with a soft voting mechanism to allow more stable and representative learning. Phase 4 depicts the Proposed Framework, which combines the two best-performing CNN and ViT models. Feature refinements were performed randomly by using Global Average Pooling and feature scaling, while a self-attention mechanism enabled the accurate cross-modal feature fusion. The generalization capability of the fused representation was further enhanced by the subsequent of dense layers followed by dropout. Results: XGBoost exhibited the highest performance among the evaluated ML classifiers, achieving 98.7% accuracy and 98.7% F1-score on BreakHis, while achieving 95.8% accuracy on external BACH dataset backed by Grad-CAM- and Grad-CAM++-based interpretability. Conclusions: By integrating CNNs and ViTs through self-attention, the proposed framework offers a robust and interpretable solution for automated breast cancer diagnosis. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

32 pages, 5551 KB

Open AccessArticle

BanglaOCT2025: A Population-Specific Fovea-Centric OCT Dataset with Self-Supervised Volumetric Restoration Using Flip-Flop Swin Transformers

by Chinmay Bepery, G. M. Atiqur Rahaman, Rameswar Debnath, Sajib Saha, Md. Shafiqul Islam, Md. Emranul Islam Abir and Sanjay Kumar Sarker

Diagnostics 2026, 16(3), 420; https://doi.org/10.3390/diagnostics16030420 - 1 Feb 2026

Viewed by 962

Abstract

Background: Age-related macular degeneration (AMD) is a major cause of vision loss, yet publicly available Optical Coherence Tomography (OCT) datasets lack demographic diversity, particularly from South Asian populations. Existing datasets largely represent Western cohorts, limiting AI generalizability. Moreover, raw OCT volumes contain redundant [...] Read more.

Background: Age-related macular degeneration (AMD) is a major cause of vision loss, yet publicly available Optical Coherence Tomography (OCT) datasets lack demographic diversity, particularly from South Asian populations. Existing datasets largely represent Western cohorts, limiting AI generalizability. Moreover, raw OCT volumes contain redundant spatial information and speckle noise, hindering efficient analysis. Methods: We introduce BanglaOCT2025, a retrospective dataset collected from the National Institute of Ophthalmology and Hospital (NIOH), Bangladesh, using Nidek RS-330 Duo 2 and RS-3000 Advance systems. We propose a novel preprocessing pipeline comprising two stages: (1) A constraint-based centroid minimization algorithm automatically localizes the foveal center and extracts a fixed 33-slice macular sub-volume, robust to retinal tilt and acquisition variability; and (2) A self-supervised volumetric denoising module based on a Flip-Flop Swin Transformer (FFSwin) backbone suppresses speckle noise without requiring paired clean reference data. Results: The dataset comprises 1585 OCT volumes (202,880 B-scans), including 857 expert-annotated cases (54 DryAMD, 61 WetAMD, and 742 NonAMD). Denoising quality was evaluated using reference-free volumetric metrics, paired statistical analysis, and blinded clinical review by a retinal specialist, confirming preservation of pathological biomarkers and absence of hallucination. Under a controlled paired evaluation using the same classifier with frozen weights, downstream AMD classification accuracy improved from 69.08% to 99.88%, interpreted as an upper-bound estimate of diagnostic signal recoverability rather than independent generalization. Conclusions: BanglaOCT2025 is the first clinically validated OCT dataset representing the Bengali population and establishes a reproducible fovea-centric volumetric preprocessing and restoration framework for AMD analysis, with future validation across independent and multi-centre test cohorts. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Graphical abstract

15 pages, 2430 KB

Open AccessArticle

Improved Detection of Small (<2 cm) Hepatocellular Carcinoma via Deep Learning-Based Synthetic CT Hepatic Arteriography: A Multi-Center External Validation Study

by Jung Won Kwak, Sung Bum Cho, Ki Choon Sim, Jeong Woo Kim, In Young Choi and Yongwon Cho

Diagnostics 2026, 16(2), 343; https://doi.org/10.3390/diagnostics16020343 - 21 Jan 2026

Viewed by 1303

Abstract

Background/Objectives: Early detection of hepatocellular carcinoma (HCC), particularly small lesions (<2 cm), which is crucial for curative treatment, remains challenging with conventional liver dynamic computed tomography (LDCT). We aimed to develop a deep learning algorithm to generate synthetic CT during hepatic arteriography (CTHA) [...] Read more.

Background/Objectives: Early detection of hepatocellular carcinoma (HCC), particularly small lesions (<2 cm), which is crucial for curative treatment, remains challenging with conventional liver dynamic computed tomography (LDCT). We aimed to develop a deep learning algorithm to generate synthetic CT during hepatic arteriography (CTHA) from non-invasive LDCT and evaluate its lesion detection performance. Methods: A cycle-consistent generative adversarial network with an attention module [Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization (U-GAT-IT)] was trained using paired LDCT and CTHA images from 277 patients. The model was validated using internal (68 patients, 139 lesions) and external sets from two independent centers (87 patients, 117 lesions). Two radiologists assessed detection performance using a 5-point scale and the detection rate. Results: Synthetic CTHA significantly improved the detection of sub-centimeter (<1 cm) HCCs compared with LDCT in the internal set (69.6% vs. 47.8%, p < 0.05). This improvement was robust in the external set; synthetic CTHA detected a greater number of small lesions than LDCT. Quantitative metrics (structural similarity index measure and peak signal-to-noise ratio) indicated high structural fidelity. Conclusions: Deep-learning–based synthetic CTHA significantly enhanced the detection of small HCCs compared with standard LDCT, offering a non-invasive alternative with high detection sensitivity, which was validated across multicentric data. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

21 pages, 2615 KB

Open AccessArticle

Evaluating the Impact of Demographic Factors on Subject-Independent EEG-Based Emotion Recognition Approaches

by Nathan Douglas, Maximilien Oosterhuis and Camilo E. Valderrama

Diagnostics 2026, 16(1), 144; https://doi.org/10.3390/diagnostics16010144 - 1 Jan 2026

Cited by 2 | Viewed by 967

Abstract

Background: Emotion recognition using electroencephalography (EEG) offers a non-invasive means of measuring brain responses to affective stimuli. However, since EEG signals can vary significantly between subjects, developing a deep learning model capable of accurately predicting emotions is challenging. Methods: To address [...] Read more.

Background: Emotion recognition using electroencephalography (EEG) offers a non-invasive means of measuring brain responses to affective stimuli. However, since EEG signals can vary significantly between subjects, developing a deep learning model capable of accurately predicting emotions is challenging. Methods: To address that challenge, this study proposes a deep learning approach that fuses EEG features with demographic information, specifically age, sex, and nationality, using an attention-based mechanism that learns to weigh each modality during classification. The method was evaluated using three benchmark datasets: SEED, SEED-FRA, and SEED-GER, which include EEG recordings of 31 subjects of different demographic backgrounds. Results: We compared a baseline model trained solely on the EEG-derived features against an extended model that fused the subjects’ EEG and demographic information. Including demographic information improved the performance, achieving 80.2%, 80.5%, and 88.8% for negative, neutral, and positive classes. The attention weights also revealed different contributions of EEG and demographic inputs, suggesting that the model learns to adapt based on subjects’ demographic information. Conclusions: These findings support integrating demographic data to enhance the performance and fairness of subject-independent EEG-based emotion recognition models. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

16 pages, 2601 KB

Open AccessArticle

Diagnostic Accuracy of an Offline CNN Framework Utilizing Multi-View Chest X-Rays for Screening 14 Co-Occurring Communicable and Non-Communicable Diseases

by Latika Giri, Pradeep Raj Regmi, Ghanshyam Gurung, Grusha Gurung, Shova Aryal, Sagar Mandal, Samyam Giri, Sahadev Chaulagain, Sandip Acharya and Muhammad Umair

Diagnostics 2026, 16(1), 66; https://doi.org/10.3390/diagnostics16010066 - 24 Dec 2025

Cited by 1 | Viewed by 1983

Abstract

Background: Chest radiography is the most widely used diagnostic imaging modality globally, yet its interpretation is hindered by a critical shortage of radiologists, especially in low- and middle-income countries (LMICs). The interpretation is both time-consuming and error-prone in high-volume settings. Artificial Intelligence (AI) [...] Read more.

Background: Chest radiography is the most widely used diagnostic imaging modality globally, yet its interpretation is hindered by a critical shortage of radiologists, especially in low- and middle-income countries (LMICs). The interpretation is both time-consuming and error-prone in high-volume settings. Artificial Intelligence (AI) systems trained on public data may lack generalizability to multi-view, real-world, local images. Deep learning tools have the potential to augment radiologists by providing real-time decision support by overcoming these. Objective: We evaluated the diagnostic accuracy of a deep learning-based convolutional neural network (CNN) trained on multi-view, hybrid (public and local datasets) for detecting thoracic abnormalities in chest radiographs of adults presenting to a tertiary hospital, operating in offline mode. Methodology: A CNN was pretrained on public datasets (Vin Big, NIH) and fine-tuned on a local dataset from a Nepalese tertiary hospital, comprising frontal (PA/AP) and lateral views from emergency, ICU, and outpatient settings. The dataset was annotated by three radiologists for 14 pathologies. Data augmentation simulated poor-quality images and artifacts. Performance was evaluated on a held-out test set (N = 522) against radiologists’ consensus, measuring AUC, sensitivity, specificity, mean average precision (mAP), and reporting time. Deployment feasibility was tested via PACS integration and standalone offline mode. Results: The CNN achieved an overall AUC of 0.86 across 14 abnormalities, with 68% sensitivity, 99% specificity, and 0.93 mAP. Colored bounding boxes improved clarity when multiple pathologies co-occurred (e.g., cardiomegaly with effusion). The system performed effectively on PA, AP, and lateral views, including poor-quality ER/ICU images. Deployment testing confirmed seamless PACS integration and offline functionality. Conclusions: The CNN trained on adult CXRs performed reliably in detecting key thoracic findings across varied clinical settings. Its robustness to image quality, integration of multiple views and visualization capabilities suggest it could serve as a useful aid for triage and diagnosis. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

16 pages, 3554 KB

Open AccessArticle

Early Detection of Cystoid Macular Edema in Retinitis Pigmentosa Using Longitudinal Deep Learning Analysis of OCT Scans

by Farhang Hosseini, Farkhondeh Asadi, Reza Rabiei, Arash Roshanpoor, Hamideh Sabbaghi, Mehrnoosh Eslami and Rayan Ebnali Harari

Diagnostics 2026, 16(1), 46; https://doi.org/10.3390/diagnostics16010046 - 23 Dec 2025

Cited by 2 | Viewed by 1429

Abstract

Background/Objectives: Retinitis pigmentosa (RP) is a progressive hereditary retinal disorder that frequently leads to vision loss, with cystoid macular edema (CME) occurring in approximately 10–50% of affected patients. Early detection of CME is crucial for timely intervention, yet most existing studies lack [...] Read more.

Background/Objectives: Retinitis pigmentosa (RP) is a progressive hereditary retinal disorder that frequently leads to vision loss, with cystoid macular edema (CME) occurring in approximately 10–50% of affected patients. Early detection of CME is crucial for timely intervention, yet most existing studies lack longitudinal data capable of capturing subtle disease progression. Methods: We propose a deep learning–based framework utilizing longitudinal optical coherence tomography (OCT) imaging for early detection of CME in patients with RP. A total of 2280 longitudinal OCT images were preprocessed using denoising and data augmentation techniques. Multiple pre-trained deep learning architectures were evaluated using a patient-wise data split to ensure robust performance assessment. Results: Among the evaluated models, ResNet-34 achieved the best performance, with an accuracy of 98.68%, specificity of 99.45%, and an F1-score of 98.36%. Conclusions: These results demonstrate the potential of longitudinal OCT–based artificial intelligence as a reliable, non-invasive screening tool for early CME detection in RP. To the best of our knowledge, this study is the first to leverage longitudinal OCT data for AI-driven CME prediction in this patient population. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

23 pages, 4282 KB

Open AccessArticle

Scaphoid Fracture Detection and Localization Using Denoising Diffusion Models

by Zhih-Cheng Huang, Tai-Hua Yang, Zhen-Li Yang and Ming-Huwi Horng

Diagnostics 2026, 16(1), 26; https://doi.org/10.3390/diagnostics16010026 - 21 Dec 2025

Cited by 1 | Viewed by 1184

Abstract

Background/Objectives: Scaphoid fractures are a common wrist injury, typically diagnosed and treated through X-ray imaging, a process that is often time-consuming. Among the various types of scaphoid fractures, occult and nondisplaced fractures pose significant diagnostic challenges due to their subtle appearance and variable [...] Read more.

Background/Objectives: Scaphoid fractures are a common wrist injury, typically diagnosed and treated through X-ray imaging, a process that is often time-consuming. Among the various types of scaphoid fractures, occult and nondisplaced fractures pose significant diagnostic challenges due to their subtle appearance and variable bone density, complicating accurate identification via X-ray images. Therefore, creating a reliable assist diagnostic system based on deep learning for the scaphoid fracture detection and localization is critical. Methods: This study proposes a scaphoid fracture detection and localization framework based on diffusion models. In Stage I, we augment the training data set by embedding fracture anomalies. Pseudofracture regions are generated on healthy scaphoid images, producing healthy and fractured data sets, forming a self-supervised learning strategy that avoids the complex and time-consuming manual annotation of medical images. In Stage II, a diffusion-based reconstruction model learns the features of healthy scaphoid images to perform high-quality reconstruction of pseudofractured scaphoid images, generating healthy scaphoid images. In Stage III, a U-Net-like network identifies differences between the target and reconstructed images, using these differences to determine whether the target scaphoid image contains a fracture. Results: After model training, we evaluated its diagnostic performance on real scaphoid images by comparing the model’s results with precise fracture locations further annotated by physicians. The proposed method achieved an image area under the receiver operating characteristic curve (AUROC) of 0.993 for scaphoid fracture detection, corresponding to an accuracy of 0.983, recall of 1.00, and precision of 0.975. For fracture localization, the model achieved a pixel AUROC of 0.978 and a pixel region overlap of 0.921. Conclusions: This study shows promise as a reliable, powerful, and scalable solution for the scaphoid fracture detection and localization. Experimental results demonstrate the strong performance of the denoising diffusion models; these models can significantly reduce diagnostic time while precisely localizing potential fracture regions, identifying conditions overlooked by the human eye. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

16 pages, 1560 KB

Open AccessArticle

Performance Comparison of U-Net and Its Variants for Carotid Intima–Media Segmentation in Ultrasound Images

by Seungju Jeong, Minjeong Park, Sumin Jeong and Dong Chan Park

Diagnostics 2026, 16(1), 2; https://doi.org/10.3390/diagnostics16010002 - 19 Dec 2025

Cited by 2 | Viewed by 2019

Abstract

Background/Objectives: This study systematically compared the performance of U-Net and variants for automatic analysis of carotid intima-media thickness (CIMT) in ultrasound images, focusing on segmentation accuracy and real-time efficiency. Methods: Ten models were trained and evaluated using a publicly available Carotid [...] Read more.

Background/Objectives: This study systematically compared the performance of U-Net and variants for automatic analysis of carotid intima-media thickness (CIMT) in ultrasound images, focusing on segmentation accuracy and real-time efficiency. Methods: Ten models were trained and evaluated using a publicly available Carotid Ultrasound Boundary Study (CUBS) dataset (2176 images from 1088 subjects). Images were preprocessed using histogram-based smoothing and resized to a resolution of 256 × 256 pixels. Model training was conducted using identical hyperparameters (50 epochs, batch size 8, Adam optimizer with a learning rate of 1 × 10⁻⁴, and binary cross-entropy loss). Segmentation accuracy was assessed using Dice, Intersection over Union (IoU), Precision, Recall, and Accuracy metrics, while real-time performance was evaluated based on training/inference times and the model parameter counts. Results: All models achieved high accuracy, with Dice/IoU scores above 0.80/0.67. Attention U-Net achieved the highest segmentation accuracy, while UNeXt demonstrated the fastest training/inference speeds (approximately 420,000 parameters). Qualitatively, UNet++ produced smooth and natural boundaries, highlighting its strength in boundary reconstruction. Additionally, the relationship between the model parameter count and Dice performance was visualized to illustrate the tradeoff between accuracy and efficiency. Conclusions: This study provides a quantitative/qualitative evaluation of the accuracy, efficiency, and boundary reconstruction characteristics of U-Net-based models for CIMT segmentation, offering guidance for model selection according to clinical requirements (accuracy vs. real-time performance). Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

13 pages, 457 KB

Open AccessArticle

LivSCP: Improving Liver Fibrosis Classification Through Supervised Contrastive Pretraining

by Yogita Dubey, Aditya Bhongade and Punit Fulzele

Diagnostics 2025, 15(24), 3226; https://doi.org/10.3390/diagnostics15243226 - 17 Dec 2025

Cited by 2 | Viewed by 1156

Abstract

Background: Deep learning models have been used in the past for non-invasive liver fibrosis classification based on liver ultrasound scans. After numerous improvements in the network architectures, optimizers, and development of hybrid methods, the performance of these models has barely improved. This [...] Read more.

Background: Deep learning models have been used in the past for non-invasive liver fibrosis classification based on liver ultrasound scans. After numerous improvements in the network architectures, optimizers, and development of hybrid methods, the performance of these models has barely improved. This creates a need for a sophisticated method that helps improve this slow-improving performance. Methods: We propose LivSCP, a method to train liver fibrosis classification models for better accuracy than the traditional supervised learning (SL). Our method needs no changes in the network architecture, optimizer, etc. Results: The proposed method achieves state-of-the-art performance, with an accuracy, precision, recall, and F1-score of 98.10% each, and an AUROC of 0.9972. A major advantage of LivSCP is that it does not require any modification to the network architecture. Our method is particularly well-suited for scenarios with limited labeled data and computational resources. Conclusions: In this work, we successfully propose a training method for liver fibrosis classification models in low-data and computation settings. By comparing the proposed method with our baseline (Vision Transformer with SL) and multiple models, we demonstrate the state-of-the-art performance of our method. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

15 pages, 3077 KB

Open AccessArticle

Integrating Numerical Data with AI-Based Image Processing Techniques to Improve the Diagnostic Accuracy of Detecting Dental Caries in Panoramic Radiographs

by Bengü Başarı, Nuran Ulusoy and Kamil Dimililer

Diagnostics 2025, 15(24), 3167; https://doi.org/10.3390/diagnostics15243167 - 12 Dec 2025

Cited by 1 | Viewed by 1164

Abstract

Background/Objectives: Dental caries is among the most common oral health problems, resulting from demineralization of dental hard tissues in acidic environments. Early diagnosis is essential to prevent severe tissue destruction, systemic complications and costly treatments. Conventional visual interpretation of panoramic radiographs, though [...] Read more.

Background/Objectives: Dental caries is among the most common oral health problems, resulting from demineralization of dental hard tissues in acidic environments. Early diagnosis is essential to prevent severe tissue destruction, systemic complications and costly treatments. Conventional visual interpretation of panoramic radiographs, though widely used, remains subjective and variable. This study evaluated the effectiveness of image processing techniques and artificial intelligence (AI)-assisted models for automated detection and classification of dental caries on panoramic radiographs, emphasizing numerical image data analysis. Methods: From 1084 panoramic radiographs, 405 were selected and classified into interproximal, occlusal and secondary caries groups. Each was segmented and one representative region was analyzed using the image data representation method. Numerical descriptors—brightness, contrast, entropy and histogram parameters—were extracted and evaluated with several machine learning algorithms. Results: Among tested models, the Decision Tree algorithm achieved the highest classification accuracy (0.988 at the 0.2 train-test ratio), showing superior and consistent results across caries types. Random Forest also demonstrated strong performance with limited training data, while Gaussian Naïve Bayes, KNN and RBFNN were less effective. Conclusions: The integration of numerical image features with AI-based models demonstrated high diagnostic accuracy and clinical interpretability, particularly with Decision Tree algorithm. These results highlight the potential of AI-assisted analysis of panoramic radiographs to enhance diagnostic reliability, reduce subjectivity and support more effective treatment planning. Further multicentre studies with larger and more diverse datasets are recommended to validate generalizability. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

23 pages, 2303 KB

Open AccessArticle

Explainable Deep Learning for Breast Lesion Classification in Digital and Contrast-Enhanced Mammography

by Samara Acosta-Jiménez, Miguel M. Mendoza-Mendoza, Carlos E. Galván-Tejada, José M. Celaya-Padilla, Jorge I. Galván-Tejada and Manuel A. Soto-Murillo

Diagnostics 2025, 15(24), 3143; https://doi.org/10.3390/diagnostics15243143 - 10 Dec 2025

Cited by 3 | Viewed by 1290

Abstract

Background: Artificial intelligence (AI) emerges as a powerful tool to assist breast cancer screening; however, its integration into different mammographic modalities remains insufficiently explored. Digital Mammography (DM) is widely accessible but presents limitations in dense breast tissue, whereas Contrast-Enhanced Spectral Mammography (CESM) [...] Read more.

Background: Artificial intelligence (AI) emerges as a powerful tool to assist breast cancer screening; however, its integration into different mammographic modalities remains insufficiently explored. Digital Mammography (DM) is widely accessible but presents limitations in dense breast tissue, whereas Contrast-Enhanced Spectral Mammography (CESM) provides functional information that enhances lesion visualization. Understanding how deep learning models behave across these modalities, and determining whether their decision-making patterns remain consistent, is essential for equitable clinical adoption. Methods: This study evaluates three convolutional neural network (CNN) architectures, ResNet-18, DenseNet-121, and EfficientNet-B0, for binary classification of breast lesions using DM and CESM images from the public CDD-CESM dataset (2006 images, three diagnostic classes). The models are trained separately on DM and CESM using three classification tasks: Normal vs. Benign, Benign vs. Malignant, and Normal vs. Malignant. A 3-fold cross-validation scheme and an independent test set are employed. Training uses transfer learning with ImageNet weights, weighted binary cross-entropy (BCE) loss, and SHapley Additive exPlanations (SHAP) analysis to visualize pixel-level relevance of model decisions. Results: CESM yields higher performance in the Normal vs. Benign and Benign vs. Malignant tasks, whereas DM achieves the highest discriminative ability in the Normal vs. Malignant comparison (EfficientNet-B0: AUC = 97%, Accuracy = 93.15%), surpassing the corresponding CESM results (AUC = 93%, Accuracy = 85.66%). SHAP attribution maps reveal anatomically coherent decision patterns in both modalities, with CESM producing sharper and more localized relevance regions due to contrast uptake, while DM exhibits broader yet spatially aligned attention. Across architectures, EfficientNet-B0 demonstrates the most stable performance and interpretability. Conclusions: CESM enhances subtle lesion discrimination through functional contrast, whereas DM, despite its simpler acquisition and wider availability, provides highly accurate and explainable outcomes when combined with modern CNNs. The consistent SHAP-based relevance observed across modalities indicates that both preserve clinically meaningful information. To the best of our knowledge, this study is the first to directly compare DM and CESM under identical preprocessing, training, and evaluation conditions using explainable deep learning models. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

16 pages, 2076 KB

Open AccessArticle

Mortality Prediction from Patient’s First Day PAAC Radiograph in Internal Medicine Intensive Care Unit Using Artificial Intelligence Methods

by Orhan Gok, Türker Fedai Cavus, Ahmed Cihad Genc, Selcuk Yaylaci and Lacin Tatli Ayhan

Diagnostics 2025, 15(24), 3138; https://doi.org/10.3390/diagnostics15243138 - 10 Dec 2025

Cited by 1 | Viewed by 744

Abstract

Introduction: This study aims to predict mortality using chest radiographs obtained on the first day of intensive care admission, thereby contributing to better planning of doctors’ treatment strategies and more efficient use of limited resources through early and accurate predictions. Methods: We retrospectively [...] Read more.

Introduction: This study aims to predict mortality using chest radiographs obtained on the first day of intensive care admission, thereby contributing to better planning of doctors’ treatment strategies and more efficient use of limited resources through early and accurate predictions. Methods: We retrospectively analyzed 510 ICU patients. After data augmentation, a total of 3019 chest radiographs were used for model training and validation, while an independent, non-augmented test set of 100 patients (100 images) was reserved for final evaluation. Seventy-four (74) radiomic features were extracted from the images and analyzed using machine learning algorithms. Model performances were evaluated using the area under the ROC curve (AUC), sensitivity, and specificity metrics. Results: A total of 3019 data samples were included in the study. Through feature selection methods, the initial 74 features were gradually reduced to 10. The Subspace KNN algorithm demonstrated the highest prediction accuracy, achieving AUC 0.88, sensitivity 0.80, and specificity 0.87. Conclusions: Machine learning algorithms such as Subspace KNN and features obtained from PAAC radiographs, such as GLCM Contrast, Kurtosis, Cobb angle, Haralick, Bilateral Infiltrates, Cardiomegaly, Skewness, Unilateral Effusion, Median Intensity, and Intensity Range, are promising tools for mortality prediction in patients hospitalized in the internal medicine intensive care unit. These tools can be integrated into clinical decision support systems to provide benefits in patient management. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

15 pages, 1875 KB

Open AccessArticle

MS-Detector: A Hierarchical Deep Learning Method to Detect Muscle Strain Using Bilateral Symmetric Ultrasound Images of the Body

by Le Zhu, Yifu Xiong, Huachao Wu, Li Zhu, Zihan Tang, Wenbin Pei, Jing Zhou and Zhidong Xue

Diagnostics 2025, 15(23), 3087; https://doi.org/10.3390/diagnostics15233087 - 4 Dec 2025

Viewed by 1035

Abstract

Background/Objectives: Muscle strain impairs mobility and quality of life, yet ultrasound diagnosis remains dependent on subjective expert interpretation, which can lead to variability in lesion detection. This study aimed to develop and evaluate MS-detector, a symmetry-aware, two-stage deep learning model that leverages bilateral [...] Read more.

Background/Objectives: Muscle strain impairs mobility and quality of life, yet ultrasound diagnosis remains dependent on subjective expert interpretation, which can lead to variability in lesion detection. This study aimed to develop and evaluate MS-detector, a symmetry-aware, two-stage deep learning model that leverages bilateral B-mode ultrasound images to automatically detect muscle strain and provide clinicians with a consistent second-reader decision-support tool in routine practice. Methods: A YOLOv5-based detector proposes candidate regions, and a Siamese convolutional neural network (CNN) compares contralateral regions to filter false positives. The dataset comprised 559 bilateral pairs from 86 patients with consensus labels. All splits were enforced at the patient level. A fixed, independent hold-out test set of 32 pairs was never used for training, tuning, or threshold selection. Five-fold cross-validation (CV) on the remaining development set was used for model selection. The operating point was pre-specified at T1 = 0.01 and T2 = 0.20. Results: The detector achieved mAP = 0.4006 (five-fold CV mean). On the hold-out set at the pre-specified operating point, MS-detector attained recall = 0.826 and precision = 0.486, improving F1/F2 over the YOLOv5 baseline by increasing precision with an acceptable recall trade-off. A representative figure illustrates the reduction in low-confidence false positives after filtering; this example is illustrative rather than aggregate. Conclusions: Leveraging contralateral symmetry in a hierarchical scheme improves detection precision while maintaining clinically acceptable recall, supporting MS-detector as a decision-support tool. Future work will evaluate generalizability across scanners and centers and assess calibrated probabilistic fusion and lesion grading. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

13 pages, 13857 KB

Open AccessArticle

Logarithmic Scaling of Loss Functions for Enhanced Self-Supervised Accelerated MRI Reconstruction

by Jaejin Cho

Diagnostics 2025, 15(23), 2993; https://doi.org/10.3390/diagnostics15232993 - 25 Nov 2025

Cited by 1 | Viewed by 871 | Correction

Abstract

Background/Objectives: Magnetic resonance imaging (MRI) is a widely used non-invasive imaging modality that provides high-fidelity soft-tissue contrast without ionizing radiation. However, acquiring high-resolution MRI scans is time-consuming, necessitating accelerated acquisition and reconstruction methods. Recently, self-supervised learning approaches have been introduced for reconstructing undersampled [...] Read more.

Background/Objectives: Magnetic resonance imaging (MRI) is a widely used non-invasive imaging modality that provides high-fidelity soft-tissue contrast without ionizing radiation. However, acquiring high-resolution MRI scans is time-consuming, necessitating accelerated acquisition and reconstruction methods. Recently, self-supervised learning approaches have been introduced for reconstructing undersampled MRI data without external fully sampled ground truth. Methods: In this work, we propose a logarithmic scaled scheme for conventional loss functions (e.g.,

ℓ_{1}

,

ℓ_{2}

) to enhance self-supervised MRI reconstruction. Standard self-supervised methods typically compute loss in the k-space domain, which tends to overemphasize low spatial frequencies while under-representing high-frequency information. Our method introduces a logarithmic scaling to adaptively rescale residuals, emphasizing high-frequency contributions and improving perceptual quality. Results: Experiments on public datasets demonstrate consistent quantitative improvements when the proposed log-scaled loss is applied within a self-supervised MRI reconstruction framework. Conclusions: The proposed approach improves reconstruction fidelity and perceptual quality while remaining lightweight, architecture-agnostic, and readily integrable into existing self-supervised MRI reconstruction pipelines. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

23 pages, 8309 KB

Open AccessArticle

Hybrid Faster R-CNN for Tooth Numbering in Periapical Radiographs Based on Fédération Dentaire Internationale System

by Yong-Shao Su, I Elizabeth Cha, Yi-Cheng Mao, Li-Hsin Chang, Zi-Chun Kao, Shun-Yuan Tien, Yuan-Jin Lin, Shih-Lun Chen, Kuo-Chen Li and Patricia Angela R. Abu

Diagnostics 2025, 15(22), 2900; https://doi.org/10.3390/diagnostics15222900 - 15 Nov 2025

Cited by 2 | Viewed by 1742

Abstract

Background/Objectives: Tooth numbering is essential because it allows dental clinicians to identify lesion locations during diagnosis, typically using the Fédération Dentaire Internationale system. However, accurate tooth numbering is challenging due to variations in periapical radiograph (PA) angles. In this study, we aimed to [...] Read more.

Background/Objectives: Tooth numbering is essential because it allows dental clinicians to identify lesion locations during diagnosis, typically using the Fédération Dentaire Internationale system. However, accurate tooth numbering is challenging due to variations in periapical radiograph (PA) angles. In this study, we aimed to develop a deep learning-based tool to assist dentists in accurately identifying teeth via tooth numbering and improve diagnostic efficiency and accuracy. Methods: We developed a Hybrid Faster Region-based Convolutional Neural Network (R-CNN) technique and a custom loss function tailored for PA tooth numbering to accelerate training. Additionally, we developed a tooth-numbering position auxiliary localization algorithm to address challenges associated with missing teeth and extensive crown loss in existing datasets. Results: We achieved a maximum precision of 95.16% utilizing the transformer-based NextViT-Faster R-CNN hybrid model, along with an accuracy increase of at least 8.5% and a 19.8% reduction in training time compared to models without the proposed tooth-numbering position auxiliary localization algorithm and conventional methods. Conclusions: The results demonstrate the effectiveness of the proposed method in overcoming challenges in PA tooth numbering within AI-assisted dental diagnostics, enhancing clinical efficiency, and reducing the risk of misdiagnosis in dental practices. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

20 pages, 6268 KB

Open AccessArticle

Automated Implant Placement Pathway from Dental Panoramic Radiographs Using Deep Learning for Preliminary Clinical Assistance

by Pei-Yi Wu, Shih-Lun Chen, Yi-Cheng Mao, Yuan-Jin Lin, Pin-Yu Lu, Kai-Hsun Yu, Kuo-Chen Li, Tsun-Kuang Chi, Tsung-Yi Chen and Patricia Angela R. Abu

Diagnostics 2025, 15(20), 2598; https://doi.org/10.3390/diagnostics15202598 - 15 Oct 2025

Cited by 3 | Viewed by 2887

Abstract

Background/Objective: Dental implant therapy requires clinicians to identify edentulous regions and adjacent teeth accurately to ensure precise and efficient implant placement. However, this process is time-consuming and subject to operator bias. To address this challenge, this study proposes an AI-assisted detection framework that [...] Read more.

Background/Objective: Dental implant therapy requires clinicians to identify edentulous regions and adjacent teeth accurately to ensure precise and efficient implant placement. However, this process is time-consuming and subject to operator bias. To address this challenge, this study proposes an AI-assisted detection framework that integrates deep learning and image processing techniques to predict implant placement pathways on dental panoramic radiographs, supporting clinical decision-making. Methods: The proposed framework is first applied to YOLO models to detect edentulous regions and employs image enhancement techniques to improve image quality. Subsequently, YOLO-OBB is utilized to extract pixel-level positional information about neighboring healthy teeth. An implant pathway orientation visualization algorithm is applied to derive clinically relevant implant placement recommendations. Results: Experimental evaluation using YOLOv9m and YOLOv8n-OBB demonstrated stable performance in both recognition and accuracy. The models achieved Precision values of 88.86% and 89.82%, respectively, with an average angular error of only 1.537° compared to clinical implant pathways annotated by dentists. Conclusions: This study presents the first AI-assisted diagnostic framework for DPR-based implant pathway prediction. The results indicate strong consistency with clinical planning, confirming its potential to enhance diagnostic accuracy and provide reliable decision support in implant dentistry. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

23 pages, 1668 KB

Open AccessArticle

Brain Stroke Classification Using CT Scans with Transformer-Based Models and Explainable AI

by Shomukh Qari and Maha A. Thafar

Diagnostics 2025, 15(19), 2486; https://doi.org/10.3390/diagnostics15192486 - 29 Sep 2025

Cited by 16 | Viewed by 4925

Abstract

Background & Objective: Stroke remains a leading cause of mortality and long-term disability worldwide, demanding rapid and accurate diagnosis to improve patient outcomes. Computed tomography (CT) scans are widely used in emergency settings due to their speed, availability, and cost-effectiveness. This study proposes [...] Read more.

Background & Objective: Stroke remains a leading cause of mortality and long-term disability worldwide, demanding rapid and accurate diagnosis to improve patient outcomes. Computed tomography (CT) scans are widely used in emergency settings due to their speed, availability, and cost-effectiveness. This study proposes an artificial intelligence (AI)-based framework for multiclass stroke classification (ischemic, hemorrhagic, and no stroke) using CT scan images from the Ministry of Health of the Republic of Turkey. Methods: We adopted MaxViT, a state-of-the-art Vision Transformer (ViT)-based architecture, as the primary deep learning model for stroke classification. Additional transformer variants, including Vision Transformer (ViT), Transformer-in-Transformer (TNT), and ConvNeXt, were evaluated for comparison. To improve model generalization and handle class imbalance, classical data augmentation techniques were applied. Furthermore, explainable AI (XAI) was integrated using Grad-CAM++ to provide visual insights into model decisions. Results: The MaxViT model with augmentation achieved the highest performance, reaching an accuracy and F1-score of 98.00%, outperforming the baseline Vision Transformer and other evaluated models. Grad-CAM++ visualizations confirmed that the proposed framework effectively identified stroke-related regions, enhancing transparency and clinical trust. Conclusions: This research contributes to the development of a trustworthy AI-assisted diagnostic tool for stroke, facilitating its integration into clinical practice and improving access to timely and optimal stroke diagnosis in emergency departments. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

21 pages, 2716 KB

Open AccessArticle

An Explainable Deep Learning Framework for Multimodal Autism Diagnosis Using XAI GAMI-Net and Hypernetworks

by Wajeeha Malik, Muhammad Abuzar Fahiem, Tayyaba Farhat, Runna Alghazo, Awais Mahmood and Mousa Alhajlah

Diagnostics 2025, 15(17), 2232; https://doi.org/10.3390/diagnostics15172232 - 3 Sep 2025

Cited by 11 | Viewed by 3476

Abstract

Background: Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by heterogeneous behavioral and neurological patterns, complicating timely and accurate diagnosis. Behavioral datasets are commonly used to diagnose ASD. In clinical practice, it is difficult to identify ASD because of the complexity of [...] Read more.

Background: Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by heterogeneous behavioral and neurological patterns, complicating timely and accurate diagnosis. Behavioral datasets are commonly used to diagnose ASD. In clinical practice, it is difficult to identify ASD because of the complexity of the behavioral symptoms, overlap of neurological disorders, and individual heterogeneity. Correct and timely identification is dependent on the presence of skilled professionals to perform thorough neurological examinations. Nevertheless, with developments in deep learning techniques, the diagnostic process can be significantly improved by automatically identifying and automatically classifying patterns of ASD-related behaviors and neuroimaging features. Method: This study introduces a novel multimodal diagnostic paradigm that combines structured behavioral phenotypes and structural magnetic resonance imaging (sMRI) into an interpretable and personalized framework. A Generalized Additive Model with Interactions (GAMI-Net) is used to process behavioral data for transparent embedding of clinical phenotypes. Structural brain characteristics are extracted via a hybrid CNN–GNN model, which retains voxel-level patterns and region-based connectivity through the Harvard–Oxford atlas. The embeddings are then fused using an Autoencoder, compressing cross-modal data into a common latent space. A Hyper Network-based MLP classifier produces subject-specific weights to make the final classification. Results: On the held-out test set drawn from the ABIDE-I dataset, a 20% split with about 247 subjects, the constructed system achieved an accuracy of 99.40%, precision of 100%, recall of 98.84%, an F1-score of 99.42%, and an ROC-AUC of 99.99%. For another test of generalizability, five-fold stratified cross-validation on the entire dataset yielded a mean accuracy of 98.56%, an F1-score of 98.61%, precision of 98.13%, recall of 99.12%, and an ROC-AUC of 99.62%. Conclusions: These results suggest that interpretable and personalized multimodal fusion can be useful in aiding practitioners in performing effective and accurate ASD diagnosis. Nevertheless, as the test was performed on stratified cross-validation and a single held-out split, future research should seek to validate the framework on larger, multi-site datasets and different partitioning schemes to guarantee robustness over heterogeneous populations. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

24 pages, 2282 KB

Open AccessArticle

Top-k Bottom All but σ Loss Strategy for Medical Image Segmentation

by Corneliu Florea, Laura Florea and Constantin Vertan

Diagnostics 2025, 15(17), 2189; https://doi.org/10.3390/diagnostics15172189 - 29 Aug 2025

Cited by 1 | Viewed by 1960

Abstract

Background/Objectives In this study we approach the problem of medical image segmentation by introducing a new loss function envelope that is derived from the Top-k loss strategy. We exploit the fact that, for semantic segmentation, the training loss is computed at two levels, [...] Read more.

Background/Objectives In this study we approach the problem of medical image segmentation by introducing a new loss function envelope that is derived from the Top-k loss strategy. We exploit the fact that, for semantic segmentation, the training loss is computed at two levels, more specifically at pixel level and at image level. Quite often, the envisaged problem has particularities that include noisy annotation at pixel level and limited data, but with accurate annotations at image level. Methods To address the mentioned issues, the Top-k strategy at image level and respectively the “Bottom all but

σ

” strategy at pixel level are assumed. To deal with the discontinuities of the differentials faced in the automatic learning, a derivative smoothing procedure is introduced. Results The method is thoroughly and successfully tested (in conjunction with a variety of backbone models) for several medical image segmentation tasks performed onto a variety of image acquisition types and human body regions. We present the burned skin area segmentation in standard color images, the segmentation of fetal abdominal structures in ultrasound images and ventricles and myocardium segmentation in cardiac MRI images, in all cases yielding performance improvements. Conclusions The proposed novel mechanism enhances model training by selectively emphasizing certain loss values by the use of two complementary strategies. The major benefits of the approach are clear in challenging scenarios, where the segmentation problem is inherently difficult or where the quality of pixel-level annotations is degraded by noise or inconsistencies. The proposed approach performs equally well in both convolutional neural networks (CNNs) and vision transformer (ViT) architectures. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

22 pages, 1882 KB

Open AccessArticle

Optimizing CNN-Based Diagnosis of Knee Osteoarthritis: Enhancing Model Accuracy with CleanLab Relabeling

by Thomures Momenpour and Arafat Abu Mallouh

Diagnostics 2025, 15(11), 1332; https://doi.org/10.3390/diagnostics15111332 - 26 May 2025

Cited by 9 | Viewed by 4486

Abstract

Background: Knee Osteoarthritis (KOA) is a prevalent and debilitating joint disorder that significantly impacts quality of life, particularly in aging populations. Accurate and consistent classification of KOA severity, typically using the Kellgren-Lawrence (KL) grading system, is crucial for effective diagnosis, treatment planning, and [...] Read more.

Background: Knee Osteoarthritis (KOA) is a prevalent and debilitating joint disorder that significantly impacts quality of life, particularly in aging populations. Accurate and consistent classification of KOA severity, typically using the Kellgren-Lawrence (KL) grading system, is crucial for effective diagnosis, treatment planning, and monitoring disease progression. However, traditional KL grading is known for its inherent subjectivity and inter-rater variability, which underscores the pressing need for objective, automated, and reliable classification methods. Methods: This study investigates the performance of an EfficientNetB5 deep learning model, enhanced with transfer learning from the ImageNet dataset, for the task of classifying KOA severity into five distinct KL grades (0–4). We utilized a publicly available Kaggle dataset comprising 9786 knee X-ray images. A key aspect of our methodology was a comprehensive data-centric preprocessing pipeline, which involved an initial phase of outlier removal to reduce noise, followed by systematic label correction using the Cleanlab framework to identify and rectify potential inconsistencies within the original dataset labels. Results: The final EfficientNetB5 model, trained on the preprocessed and Cleanlab-remediated data, achieved an overall accuracy of 82.07% on the test set. This performance represents a significant improvement over previously reported benchmarks for five-class KOA classification on this dataset, such as ResNet-101 which achieved 69% accuracy. The substantial enhancement in model performance is primarily attributed to Cleanlab’s robust ability to detect and correct mislabeled instances, thereby improving the overall quality and reliability of the training data and enabling the model to better learn and capture complex radiographic patterns associated with KOA. Class-wise performance analysis indicated strong differentiation between healthy (KL Grade 0) and severe (KL Grade 4) cases. However, the “Doubtful” (KL Grade 1) class presented ongoing challenges, exhibiting lower recall and precision compared to other grades. When evaluated against other architectures like MobileNetV3 and Xception for multi-class tasks, our EfficientNetB5 demonstrated highly competitive results. Conclusions: The integration of an EfficientNetB5 model with a rigorous data-centric preprocessing approach, particularly Cleanlab-based label correction and outlier removal, provides a robust and significantly more accurate method for five-class KOA severity classification. While limitations in handling inherently ambiguous cases (such as KL Grade 1) and the small sample size for severe KOA warrant further investigation, this study demonstrates a promising pathway to enhance diagnostic precision. The developed pipeline shows considerable potential for future clinical applications, aiding in more objective and reliable KOA assessment. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

20 pages, 1989 KB

Open AccessArticle

Deep Reinforcement Learning for CT-Based Non-Invasive Prediction of SOX9 Expression in Hepatocellular Carcinoma

by Minghui Liu, Yi Wei, Tianshu Xie, Meiyi Yang, Xuan Cheng, Lifeng Xu, Qian Li, Feng Che, Qing Xu, Bin Song and Ming Liu

Diagnostics 2025, 15(10), 1255; https://doi.org/10.3390/diagnostics15101255 - 15 May 2025

Cited by 1 | Viewed by 1233

Abstract

Background: The transcription factor SOX9 plays a critical role in various diseases, including hepatocellular carcinoma (HCC), and has been implicated in resistance to sorafenib treatment. Accurate assessment of SOX9 expression is important for guiding personalized therapy in HCC patients; however, a reliable non-invasive [...] Read more.

Background: The transcription factor SOX9 plays a critical role in various diseases, including hepatocellular carcinoma (HCC), and has been implicated in resistance to sorafenib treatment. Accurate assessment of SOX9 expression is important for guiding personalized therapy in HCC patients; however, a reliable non-invasive method for evaluating SOX9 status remains lacking. This study aims to develop a deep learning (DL) model capable of preoperatively and non-invasively predicting SOX9 expression from CT images in HCC patients. Methods: We retrospectively analyzed a dataset comprising 4011 CT images from 101 HCC patients who underwent surgical resection followed by sorafenib therapy at West China Hospital, Sichuan University. A deep reinforcement learning (DRL) approach was proposed to enhance prediction accuracy by identifying and focusing on image regions highly correlated with SOX9 expression, thereby reducing the impact of background noise. Results: Our DRL-based model achieved an area under the curve (AUC) of 91.00% (95% confidence interval: 88.64–93.15%), outperforming conventional DL methods by over 10%. Furthermore, survival analysis revealed that patients with SOX9-positive tumors had significantly shorter recurrence-free survival (RFS) and overall survival (OS) compared to SOX9-negative patients, highlighting the prognostic value of SOX9 status. Conclusions: This study demonstrates that a DRL-enhanced DL model can accurately and non-invasively predict SOX9 expression in HCC patients using preoperative CT images. These findings support the clinical utility of imaging-based SOX9 assessment in informing treatment strategies and prognostic evaluation for patients with advanced HCC. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

Review

Jump to: Research, Other

36 pages, 1319 KB

Open AccessReview

A Review of U-Net Based Deep Learning Frameworks for MRI-Based Brain Tumor Segmentation

by Ayse Bastug Koc and Devrim Akgun

Diagnostics 2026, 16(4), 506; https://doi.org/10.3390/diagnostics16040506 - 7 Feb 2026

Cited by 7 | Viewed by 1955

Abstract

Automated segmentation of brain tumors from Magnetic Resonance Imaging (MRI) images is helpful for clinical diagnosis, surgical planning, and post-treatment monitoring. In recent years, the U-Net architecture has been observed as one of the most popular solutions among deep learning models. This article [...] Read more.

Automated segmentation of brain tumors from Magnetic Resonance Imaging (MRI) images is helpful for clinical diagnosis, surgical planning, and post-treatment monitoring. In recent years, the U-Net architecture has been observed as one of the most popular solutions among deep learning models. This article presents a review of 35 studies published between 2019 and 2025 focusing on U-Net-based brain tumor segmentation. The primary focus of this review is an in-depth analysis of commonly used U-Net architectures. The transformation of original 2D and 3D models into more advanced variants is examined in detail. Results from a wide range of studies are synthesized, and standard evaluation criteria are summarized along with benchmark datasets such as the BRATS competition to validate the effectiveness of these models. Additionally, the paper overviews the recent developments in the field, determines fundamental challenges, and provides insight into future directions, including improving model efficiency and generalization, combining multimodal data, and advancing clinical applications. This review serves as a guide for researchers to examine the impact of the U-Net architecture on brain tumor segmentation. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

Other

Jump to: Research, Review

33 pages, 1398 KB

Open AccessSystematic Review

A Systematic Review on Synthetic Medical Images Generation—Recent Trends and Future Opportunities

by Yumna Waheed, Muhammad Nouman Noor and Imran Ashraf

Diagnostics 2026, 16(14), 2166; https://doi.org/10.3390/diagnostics16142166 - 10 Jul 2026

Viewed by 1075

Abstract

Background/Objectives: Generative Models have revolutionized the synthesis of complex, realistic medical images. However, obtaining annotated, high-quality datasets is challenging and expensive due to privacy concerns, high human annotation costs, and data scarcity. This systematic literature review (SLR) provides a comparative overview of Generative [...] Read more.

Background/Objectives: Generative Models have revolutionized the synthesis of complex, realistic medical images. However, obtaining annotated, high-quality datasets is challenging and expensive due to privacy concerns, high human annotation costs, and data scarcity. This systematic literature review (SLR) provides a comparative overview of Generative AI models used for medical image generation, categorizing the research into GANs, Diffusion Models, and Autoencoders used for medical image generation and GAN loss functions. Methods: This systematic review followed the PRISMA 2020 guidelines. Studies published between 2021 and 2026 were retrieved from selected scientific databases and screened using predefined inclusion criteria. We selected 120 papers published in reputable databases and journals. The review examined generative modeling approaches aimed at addressing data scarcity and annotation limitations in medical imaging, including GAN-based adversarial synthesis, diffusion-based iterative denoising, and Autoencoder-based latent representation learning, along with an in-depth analysis of GAN loss functions. Results: GANs were the most widely used approach (36% of the reviewed studies), achieving the lowest FID scores (15.552–31.349). Diffusion Models (26% of the reviewed studies) showed superior structural fidelity, achieving SSIM values of up to 0.915, whereas GANs achieved 0.593 on brain MRI datasets. In addition, hybrid loss functions combining perceptual, structural, and pixel-level terms resulted in improved image quality performance (PSNR = 35.6; SSIM = 0.95). Conclusions: The analysis showed that GAN-based generative models produce visually realistic images with detailed textures, whereas Diffusion Models generate high-resolution images with superior structural fidelity but require substantial computational resources for training and image generation. Furthermore, the clinical implications, limitations, and challenges of Generative AI models were discussed. Despite these advances, the medical imaging field still faces several challenges, including the need for annotated medical datasets, high-quality realistic images for training, and limited data availability. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

1 pages, 820 KB

Open AccessCorrection

Correction: Cho, J. Logarithmic Scaling of Loss Functions for Enhanced Self-Supervised Accelerated MRI Reconstruction. Diagnostics 2025, 15, 2993

by Jaejin Cho

Diagnostics 2026, 16(4), 553; https://doi.org/10.3390/diagnostics16040553 - 13 Feb 2026

Viewed by 410

Abstract

In the original publication [...] Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 2

9 pages, 1528 KB

Open AccessBrief Report

Impact of Deep Learning-Based Reconstruction on the Accuracy and Precision of Cardiac Tissue Characterization

by Margarita Gorodezky, Linda Reichardt, Tom Geisler, Marc-André Weber, Felix G. Meinel and Ann-Christin Klemenz

Diagnostics 2026, 16(2), 348; https://doi.org/10.3390/diagnostics16020348 - 21 Jan 2026

Cited by 1 | Viewed by 1173

Abstract

Background/Objectives: Interest in myocardial mapping for cardiac MRI has increased, enabling differentiation of various cardiac diseases through T1, T2, and T2* mapping. This study evaluates the impact of deep learning (DL)-based image reconstruction on the mean and standard deviation (SD) of these techniques. [...] Read more.

Background/Objectives: Interest in myocardial mapping for cardiac MRI has increased, enabling differentiation of various cardiac diseases through T1, T2, and T2* mapping. This study evaluates the impact of deep learning (DL)-based image reconstruction on the mean and standard deviation (SD) of these techniques. Methods: Fifty healthy adults underwent cardiac MRI, with images reconstructed using the AIR Recon DL prototype. This DL approach, which reduces noise and enhances image quality, was applied at three levels and compared to non-DL reconstructions. Results: Analysis focused on the septum to minimize artifacts. For T1 mapping, mean values were 988 ± 50, 981 ± 45, 982 ± 43, and 980 ± 24 ms; for T2 mapping, mean values were 53 ± 5, 54 ± 5, 54 ± 5, and 54 ± 5 ms and for T2* mapping, mean values were 37 ± 5, 37 ± 5, 37 ± 5, and 38 ± 5 ms for no DL and increasing DL levels, respectively. Results showed no significant differences in mean values for any mappings between non-DL and DL reconstructions. However, DL significantly reduced the SD within regions of interest for T1 mapping, enhancing image sharpness and registration accuracy. No significant SD reduction was observed for T2 and T2* mappings. Conclusions: These findings suggest that DL-based reconstructions improve the precision of T1 mapping without affecting mean values, supporting their clinical integration for more accurate cardiac disease diagnosis. Future studies should include patient cohorts and optimized protocols to further validate these findings. Full article

(This article belongs to the Special Issue 3rd Edition: AI/ML-Based Medical Image Processing and Analysis)

► Show Figures

Figure 1

Journal Menu

Journal Browser

3rd Edition: AI/ML-Based Medical Image Processing and Analysis

Share This Special Issue

Editors

Special Issue Information

Keywords

Benefits of Publishing in a Special Issue

Related Special Issue

Published Papers (36 papers)

Research

Review

Other

Further Information

Guidelines

MDPI Initiatives

Follow MDPI