Explainable Two-Stage Xception-Swin Transformer Learning for Body-Part-Aware Fracture Detection in Musculoskeletal...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Automated interpretation of upper-extremity radiographs is challenging due to subtle fractures and anatomical variations.
• Class imbalance in datasets further complicates accurate fracture detection in X-rays.

Purpose of the Study:

• To develop and evaluate a two-stage deep learning framework for improved body-part recognition and abnormality detection in musculoskeletal radiographs.
• To enhance the accuracy of automated fracture interpretation in upper-extremity X-rays.

Main Methods:

• A hybrid Xception-Swin deep learning model was developed, combining local structural features with transformer-based contextual features.
• The framework employed attention-based fusion and was evaluated in a two-stage process: body-part classification followed by abnormality detection within anatomical subsets.
• Performance metrics included accuracy, F1-score, AUC-ROC, Cohen's kappa, calibration, ablation studies, and zero-shot validation on the FracAtlas dataset.

Main Results:

• The model achieved high performance in body-part classification (accuracy=0.9643, macro F1=0.9574).
• Abnormality detection accuracy ranged from 0.7289 to 0.8538, with F1 scores from 0.7191 to 0.8508.
• Hybrid attention mechanisms improved performance, and preliminary cross-dataset generalizability was demonstrated via FracAtlas validation (AUC=0.8247, kappa=0.5812).

Conclusions:

• The proposed two-stage deep learning framework effectively improves automated interpretation of upper-extremity radiographs.
• Complementary fusion of Convolutional Neural Networks (CNNs) and Transformers shows promise for enhancing diagnostic accuracy in medical imaging.
• The study indicates preliminary cross-dataset generalizability, supporting the framework's potential for wider clinical application.