International Journal of Imaging Systems and Technology最新文献

{"title":"Comparative Analysis of Single-Stream and Multi-Stream Models for Medical Image Fusion","authors":"Kwasi Odei Boadu,&nbsp;Jin Qi,&nbsp;David Ocho Audu","doi":"10.1002/ima.70453","DOIUrl":"https://doi.org/10.1002/ima.70453","url":null,"abstract":"<div>\u0000 \u0000 <p>Medical image fusion supports more informative image interpretation by integrating complementary information from different imaging modalities. This paper presents a controlled comparative study of three input-stream configurations for deep learning-based multi-modal medical image fusion on Positron Emission Tomography-Magnetic Resonance Imaging (PET-MRI) and Single-Photon Emission Computed Tomography-Magnetic Resonance Imaging (SPECT-MRI) fusion tasks. The evaluated configurations include (1) Single-Input, which fuses the two modalities at the input level; (2) Dual-Input, which processes each modality through a separate encoder branch before feature fusion; and (3) Tri-Input, which combines individual modality streams with an additional concatenated input stream to learn both modality-specific and joint representations. All configurations use the same U-Net-based encoder-decoder backbone with Transformer-CNN modules, allowing the effect of input-stream design to be examined under a consistent architectural and training setting. Performance is evaluated using entropy (EN), mutual information (MI), feature mutual information (FMI), correlation coefficient (CC), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), validation loss, training time per epoch, external baseline comparison, and statistical significance analysis. The results show that Dual-Input provides the most balanced overall performance, especially for PET-MRI fusion, while Tri-Input achieves stronger feature-transfer performance on selected SPECT-MRI metrics. Single-Input offers the fastest training speed but lower overall fusion quality among the evaluated stream configurations. These findings clarify the trade-offs among early fusion, separate modality processing, and hybrid multi-stream input design for functional-anatomical medical image fusion, while remaining limited to the evaluated dataset, modality pairs, and image-quality metrics.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-10-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149205712","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MetaBayes-SkinNet: A Lightweight Multimodal Framework With Meta-Aware Bayesian-Optimized Augmentation for Multiclass Skin Lesion Classification","authors":"Qusay Saihood,&nbsp;Emrullah Sonuç,&nbsp;Yusuf Yargı Baydilli","doi":"10.1002/ima.70459","DOIUrl":"https://doi.org/10.1002/ima.70459","url":null,"abstract":"<div>\u0000 \u0000 <p>Accurate multiclass skin lesion classification is essential for computer-aided diagnosis (CAD) systems that support early detection and clinical decision-making. Despite advances in artificial intelligence-based CAD systems, key challenges persist, including severe class imbalance, limited classification accuracy, high computational cost, and insufficient model interpretability. To address these challenges, this study proposes a lightweight, interpretable, multimodal framework for imbalanced multiclass skin lesion classification. The framework introduces a meta-aware Bayesian-optimized RandAugment strategy that applies anatomy-specific augmentation policies to address class imbalance while preserving clinically plausible variations. A lightweight multimodal MobileNetV3Small architecture integrates dermoscopic image features with clinical metadata through late fusion for efficient representation learning. The resulting multimodal features are classified using a Bayesian-optimized Light Gradient Boosting Machine (LGBM), and model interpretability is provided through complementary Grad-CAM and Integrated Gradients visualizations. Experimental evaluation on the HAM10000 dataset demonstrates that the proposed framework achieves 98.17% classification accuracy under lesion-grouped stratified five-fold cross-validation with only 1.35 M parameters and 0.114 GFLOPs, providing highly competitive predictive performance together with substantial computational efficiency. These results indicate that the proposed framework offers a practical, accurate, and interpretable solution for multiclass skin lesion classification, with potential for clinical decision support, subject to prospective external validation.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149205457","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Interpretable Skin Lesion Classification via Attention-Augmented Ensemble With Two-Stage Training","authors":"Ali Han Özdoğan,&nbsp;Gönül Sinem Özdoğan,&nbsp;Selcan Kaplan Berkaya","doi":"10.1002/ima.70458","DOIUrl":"https://doi.org/10.1002/ima.70458","url":null,"abstract":"<div>\u0000 \u0000 <p>Skin cancer, particularly, melanoma, remains a major global health challenge, with early detection being critical for effective treatment and improved patient outcomes. Traditional diagnostic approaches, which rely on clinician expertise, are susceptible to variability and have limited scalability. Deep learning models often struggle with extreme class imbalance and poor interpretability in this context. We propose an interpretable ensemble framework comprising three CNN backbones—EfficientNet-B0, DenseNet-169, and ResNet-50—each evaluated with and without a custom channel attention mechanism; the final ensemble combines EfficientNet-B0 + CA, DenseNet-169 + CA, and ResNet-50. To enhance feature discriminability for minority classes, we introduce a specialized two-stage training strategy: initial stabilization on balanced subsets followed by global fine-tuning on the full distribution. The weighted ensemble achieved 90.81% accuracy and an 86.68% <i>F</i>₁-score on the HAM10000 dataset. Grad-CAM visualizations provide post hoc evidence, confirming the model's focus on clinically relevant pathological markers. This study demonstrates that architectural refinements combined with strategic learning protocols significantly improve multiclass diagnostic performance in imbalanced medical datasets.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146417","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"FBDA-Net: Medical Image Segmentation via Foreground–Background Decoupled Attention Aggregation","authors":"Honggang Xie,&nbsp;Tong Liu,&nbsp;Mujun Zang,&nbsp;Shusen Zhou,&nbsp;Chanjuan Liu,&nbsp;Qingjun Wang","doi":"10.1002/ima.70454","DOIUrl":"https://doi.org/10.1002/ima.70454","url":null,"abstract":"<div>\u0000 \u0000 <p>Medical image segmentation is a fundamental task in computer-aided diagnosis, playing an important role in organ measurement, lesion delineation, and surgical planning. However, due to the limitations of imaging conditions and the complexity of tissue characteristics, foreground targets and the background often exhibit a “soft boundary,” and the two are easily confused in ambiguous transition regions, which limits further improvement of segmentation accuracy. To address this, we propose FBDA-Net, a segmentation network based on foreground–background decoupled attention (FBDA). Its core module, the FBDA, decouples the input features into foreground and background at the feature level using prior masks, computes attention with a shared Query and three independent sets of Key/Value, and employs a joint Softmax so that the foreground, background, and global branches competitively allocate weights within the same probability space. Building upon this, we construct a top-down progressive decoupling decoder that converts the prediction at each level into prior masks and passes them downward layer by layer, driving the decoupled priors to be progressively reconstructed and refined along the decoding process, and we adopt a combinatorial supervision loss to strengthen the cross-level consistency of multi-level predictions. Experiments on three benchmark datasets—Synapse multi-organ CT, ACDC cardiac MRI, and ISIC 2016 dermoscopy—demonstrate competitive overall segmentation performance. Zero-shot external evaluation on PH2 without fine-tuning further supports the cross-dataset generalization of FBDA-Net. The source code is publicly available at https://github.com/TongLiuLDU/FBDANet.git.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146407","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"ADW-SwinUNet for Skin Melanoma Electrical Impedance Tomography Image Reconstruction Using a Honeycomb-Structured Sensor","authors":"Chengtao Yu,&nbsp;Xinyu Zhou,&nbsp;Xiang Li,&nbsp;Huaijin Chen,&nbsp;Li Ding,&nbsp;Songpei Hu,&nbsp;Xia Ye,&nbsp;Jiafeng Yao","doi":"10.1002/ima.70446","DOIUrl":"https://doi.org/10.1002/ima.70446","url":null,"abstract":"<div>\u0000 \u0000 <p>Melanoma is a highly malignant skin tumor, and early detection with accurate boundary delineation is critical for improving patient survival. Electrical impedance tomography (EIT) is a promising technique for melanoma detection due to its noninvasiveness, low cost, and portability. However, current EIT reconstruction methods are prone to boundary blurring, structural distortion, and missing details owing to the highly ill-posed nature of the inverse problem. To address this issue, this paper develops an Adaptive Dynamic Window SwinUNet (ADW-SwinUNet) for EIT-based melanoma imaging. The acquired 16 × 13 boundary voltage vector is transformed into a 128 × 128 feature map and then fed into the network. By introducing an ADW mechanism, the network enables hierarchical multi-scale feature modeling, thereby improving the balance between global conductivity representation and local boundary recovery. In addition, a 16-electrode honeycomb-structured handheld sensor is designed to improve the sensitivity distribution within the sensing domain. Compared with the conventional ring-shaped configuration, the proposed planar electrode configuration enhances the sensitivity in the central and transition regions, thereby improving the response to internal abnormalities. Simulation results demonstrate that ADW-SwinUNet outperforms Tikhonov regularization, convolutional neural network-based methods, and DHU-Net, a high-capacity network based on an improved U-Net architecture, achieving an average root mean square error (RMSE) of 0.044 and an intraclass correlation coefficient of 0.98. Experimental results validate that the proposed method exhibits high robustness, with RMSE values ranging from 0.06 to 0.10. These results indicate the potential of the proposed method for noninvasive and portable EIT-based skin melanoma imaging.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-10-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149145929","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Similarity-Aware Evaluation, Probability Calibration, and Independent Replication for Reliable Gastrointestinal Image Classification","authors":"Housheng Liu,&nbsp;Yingwu Xu,&nbsp;Fang Cheng,&nbsp;Jinlian Zha","doi":"10.1002/ima.70452","DOIUrl":"https://doi.org/10.1002/ima.70452","url":null,"abstract":"<div>\u0000 \u0000 <p>High accuracy alone does not establish reliable medical-image classification because visually similar images may cross partitions and confidence may be miscalibrated. We evaluated a similarity-aware framework on a four-class gastrointestinal benchmark (4000 images) and independently replicated the methodology on GastroVision (8000 images; 27 classes). A 64-bit DCT perceptual hash defined groups at Hamming distance ≤ 4. Five-fold random and group-constrained protocols were matched by model and seed, with separate validation, calibration, and test subsets. We compared modern backbones, temperature scaling, a 2 × 2 Mixup/label-smoothing ablation, and deterministic maximum-softmax probability. On the primary benchmark, grouped EfficientNet-B0 achieved accuracy and macro-F1 of 0.9765 ± 0.0022. Grouping changed accuracy by +0.0025 relative to random splitting (Wilcoxon <i>p</i> = 0.625). Temperature scaling reduced expected calibration error from 0.1292 ± 0.0216 to 0.0123 ± 0.0030 without changing predictions, and deterministic MSP achieved error-detection AUROC of 0.9288 ± 0.0181. On GastroVision, random folds contained 78–101 detected cross-subset overlaps, whereas grouped folds contained none. Grouped EfficientNet-B0 achieved 0.8195 ± 0.0043 accuracy and 0.5792 ± 0.0148 macro-F1; ECE decreased from 0.1114 ± 0.0148 to 0.0323 ± 0.0048 after scaling. Similarity grouping removed detected resemblance without reducing accuracy, while calibration and uncertainty analyses exposed properties hidden by discrimination metrics. GastroVision supports methodological portability rather than direct external validation; pHash cannot establish patient or procedure independence, for which identifiers remain preferable.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149145877","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"PMCA-UNet: A Parameter-Efficient Skin Lesion Segmentation Network With Initial Feature Preservation and Parallel Multi-Scale Convolutional Attention","authors":"Zhiwei Sun,&nbsp;Yuxin Xing,&nbsp;Guanglei Zhang,&nbsp;Yuliang Liu","doi":"10.1002/ima.70455","DOIUrl":"https://doi.org/10.1002/ima.70455","url":null,"abstract":"<div>\u0000 \u0000 <p>Skin lesion segmentation is a critical component of computer-aided dermoscopic image analysis, and its accuracy directly affects subsequent evaluation reliability. However, small lesions, ambiguous boundaries, and foreground–background imbalance remain challenging, while high-performance methods often involve large model sizes that limit deployment in resource-constrained scenarios. To address these issues, this paper proposes PMCA-UNet, a parameter-efficient skin lesion segmentation network based on AMSUnet. The proposed model introduces an enhanced initial feature preservation module (E-PFC) to strengthen shallow texture and boundary information, and a parallel multi-scale convolutional attention module (P-MCA) to improve contextual representation for lesions with different scales. In addition, the CE-Dice hybrid loss weighting strategy is optimized to alleviate lesion–background pixel imbalance. On the ISIC2018 dataset, PMCA-UNet achieves Dice and IoU values of 0.9184 and 0.8642, improving the AMSUnet baseline by 3.47 and 5.47 percentage points, respectively. Its sensitivity reaches 0.9478, indicating improved lesion recall capability. The model also achieves Dice scores of 0.9256 and 0.9408 on the ISIC2016 and PH2 datasets, respectively. With only 3.186 M parameters, PMCA-UNet demonstrates a favorable balance between segmentation performance and model efficiency. The source code is available at: https://github.com/hang-hang683/PMCA-Seg.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-10-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149145724","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A Hybrid Vision Transformer Based Edge-Guided Boundary Attention U-Net for Skin Lesion Segmentation","authors":"M. Cime Monika,&nbsp;R. Baby Shalini","doi":"10.1002/ima.70449","DOIUrl":"https://doi.org/10.1002/ima.70449","url":null,"abstract":"<div>\u0000 \u0000 <p>Accurate skin lesion segmentation plays a vital role in computer-aided diagnosis for early detection and timely treatment of skin cancer. However, existing deep learning models often struggle to simultaneously capture long-range contextual dependencies and preserve fine lesion boundary details in challenging dermoscopic images. To address these limitations, we propose Trans-EGBAU-Net, a hybrid Vision Transformer (ViT) based Edge-Guided Boundary Attention (EGBA) U-Net for accurate and boundary-aware skin lesion segmentation. The proposed framework employs a ViT encoder to learn global contextual representation from image patches through multiple transformer layers. To compensate for the limited local spatial sensitivity of transformer features, a lightweight Edge Attention Module (EAM) is incorporated after the encoder to enhance edge-aware representations around lesion boundaries. Furthermore, the proposed EGBA mechanism utilizes edge-enhanced transformer features generated by the EAM to guide the Boundary-Guided Attention Gate (BGAG) during skip-feature refinement, while a standard Attention Gate (AG) selectively emphasizes semantically relevant lesion regions. The complementary integration of AG and EGBA enables simultaneous semantic feature selection and contour-aware boundary refinement, resulting in improved delineation of irregular and low-contrast lesion boundaries. Experimental evaluation on the HAM10000, ISIC 2018, and PH2 datasets demonstrates competitive segmentation performance, achieving accuracies of 0.9793, 0.9688, and 0.9679; Dice coefficients of 0.9446, 0.9165, and 0.9155; IoU scores of 0.9009, 0.8753, and 0.8596; Sensitivities of 0.9796, 0.9345, and 0.9680; and Specificities of 0.9849, 0.9772, and 0.9488, respectively. Ablation studies further validate the effectiveness of the proposed attention mechanisms, making Trans-EGBAU-Net a reliable framework for boundary-aware skin lesion segmentation.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-09-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146870","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"SurpSAM: Saliency- and Uncertainty-Guided Reinforcement Prompting for Automatic Left Atrial Segmentation","authors":"Yaodan Wang,&nbsp;Hualing Li,&nbsp;Jiaxin Huo","doi":"10.1002/ima.70434","DOIUrl":"https://doi.org/10.1002/ima.70434","url":null,"abstract":"<div>\u0000 \u0000 <p>Left atrial (LA) CT segmentation is essential for atrial fibrillation assessment, ablation treatment planning, and three-dimensional cardiac structural analysis. However, limited annotations, indistinct anatomical boundaries, and substantial inter-individual variability make accurate segmentation challenging. To address this problem, we propose SurpSAM, a saliency- and uncertainty-guided reinforcement prompt optimization framework for one-shot LA CT segmentation. Given only one annotated reference sample, SurpSAM first constructs an LA-specific structural prior through the la-specific structural prior module (SSPM), which provides anatomy-aware saliency responses on target CT slices. Based on this prior, an uncertainty-aware prompt generation strategy (UPGS) samples complementary positive and negative points to form the initial prompt set for the frozen Segment Anything Model (SAM). Rather than directly using these initial prompts, SurpSAM performs Prompt-Space Action Inference, in which a PPO-trained policy adaptively updates the prompt set according to saliency, uncertainty, and SAM feedback. To improve volumetric structural reliability, a subsequent mask action refinement module constructs multiple candidate masks in the mask space and selects the final segmentation according to saliency consistency, reference prior agreement, topological continuity, and prompt consistency. Experiments on the MM-WHS and ImageCAS-STACOM datasets show that SurpSAM achieves Dice scores of 76.55% and 83.53%, respectively, using only one annotated sample, outperforming existing one-shot segmentation methods and training-free SAM prompting strategies. These results demonstrate the effectiveness of progressive prompt generation, reinforcement-based prompt updating, and mask-space structural refinement for robust one-shot LA CT segmentation.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-09-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149068327","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multi-Modal Transformer Framework for Automated Knee Osteoarthritis Severity Grading From Radiographs","authors":"Qiang Yue,&nbsp;Qian Wang,&nbsp;Changshuang Bai,&nbsp;ZhaoXue Li,&nbsp;Li Zhao","doi":"10.1002/ima.70450","DOIUrl":"https://doi.org/10.1002/ima.70450","url":null,"abstract":"<div>\u0000 \u0000 <p>To develop and validate a multi-modal, artificial intelligence (AI)-driven framework for automated grading of knee osteoarthritis severity from radiographs, integrating advanced detection, deep learning, radiomics, and clinical data. A total of 2290 patients from five medical centers were included, split into an internal cohort (<i>n</i> = 1549) and an external test set (<i>n</i> = 741). Knee radiographs were preprocessed with histogram normalization, contrast-limited adaptive histogram equalization (CLAHE), and noise filtering. Four detection models, RT-DETR, SAM, DINOv2 + ViTDet, and Grounded-SAM/MedSAM, were evaluated for anatomical localization, with Grounded-SAM/MedSAM selected for optimal intersection over union (IoU) and mean average precision (mAP). Radiomic features (<i>n</i> = 215) were extracted from detected regions of interest (ROIs) using the Standardized Environment for Radiomics Analysis (SERA) platform, while deep features were derived from four advanced extractors, with MedSAM + feature aggregator selected for highest classification accuracy. Features underwent harmonization, variance filtering, and selection via least absolute shrinkage and selection operator (LASSO), principal component analysis (PCA), or mutual information. Clinical, radiomic, and deep features were fused and classified using CoAtNet, ConvNeXt V2, EfficientNetV2, RepLKNet, and Swin V2 Transformers. Model selection was performed within a nested cross-validation scheme on the internal cohort, and the external cohort was accessed only once after the final configuration had been fixed. Model performance was assessed by accuracy, F1-score, and area under the receiver operating characteristic curve (AUC), with interpretability provided by SHapley Additive exPlanations (SHAP) analysis and a nomogram. Grounded-SAM/MedSAM achieved the highest detection performance (external IoU = 90.83% [95% confidence interval, CI, 89.41–92.15], mAP = 93.67% [95% CI 92.38–94.82]). MedSAM deep features outperformed others (external F1 = 82.37% [95% CI 80.56–84.18], AUC = 0.902 [95% CI 0.888–0.916]). Fused features with Swin V2 + LASSO yielded the best results (external accuracy = 94.6% [95% CI 93.5–95.7], F1 = 94.4% [95% CI 93.3–95.5], AUC = 97.2% [95% CI 96.4–98.0]), significantly surpassing single-modality models. SHAP analysis confirmed balanced contributions from deep, radiomic, and clinical features, while the nomogram achieved a C-index of 0.970 [95% CI 0.958–0.980] with excellent calibration. This framework demonstrates high accuracy, robustness, and interpretability in knee osteoarthritis severity grading from radiographs, offering strong potential for real-world clinical integration.</p>\u0000 </div>","PeriodicalId":14027,"journal":{"name":"International Journal of Imaging Systems and Technology","volume":"36 6","pages":""},"PeriodicalIF":3.0,"publicationDate":"2026-09-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149146840","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}