Explainable Two-Stage Xception-Swin Transformer Learning for Body-Part-Aware Fracture Detection in Musculoskeletal X-Rays.
Authors
Affiliations (5)
Affiliations (5)
- Department of Computer Science, COMSATS University Islamabad (CUI), Wah Campus, Wah 47000, Pakistan.
- Department of Research and Development (R&D), GPI SpA, 38123 Trento, Italy.
- PolitoBIOMed Lab, Department of Mechanical and Aerospace Engineering, Politecnico di Torino, 10129 Turin, Italy.
- Centro Medico Santagostino, 20127 Milan, Italy.
- Department of Computer and Information Sciences, Pakistan Institute of Engineering and Applied Sciences (PIEAS), Islamabad 45650, Pakistan.
Abstract
Accurate automated interpretation of upper-extremity musculoskeletal radiographs remains challenging because fracture appearance varies across anatomical regions and can be subtle under class imbalance. This study proposes a two-stage deep learning framework for MURA-based X-ray analysis, aiming to improve body-part recognition and body-part-wise abnormality detection. Multiple architectures were first compared for seven-class body-part classification, after which the selected hybrid Xception-Swin model was fine-tuned for abnormality detection within each anatomical subset. The framework combines Xception-derived local structural features with Swin Transformer contextual features using attention-based fusion, and performance was evaluated using accuracy, F1-score, AUC-ROC, Cohen's kappa, calibration, component-level ablation, post hoc explainability, and zero-shot FracAtlas validation. For body-part classification, the model achieved accuracy = 0.9643, macro F1 = 0.9574, AUC-ROC = 0.9963, and kappa = 0.9579. For abnormality detection, accuracy ranged from 0.7289 to 0.8538, F1 from 0.7191 to 0.8508, AUC from 0.7693 to 0.9080, and kappa from 0.4449 to 0.7071. Ablation on hand and humerus radiographs showed the highest macro F1 with Hybrid Attention, while FracAtlas validation yielded AUC = 0.8247 and kappa = 0.5812. The results support complementary CNN-Transformer fusion and indicate preliminary cross-dataset generalizability. Implementation resources are available at Zenodo.