Interpretable deep learning for rotator cuff calcific tendinopathy diagnosis: a multi-center study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiography

Background:

• Reliable artificial intelligence (AI) deployment in medical imaging necessitates high diagnostic performance, robustness, and interpretability.
• Automated analysis of shoulder radiographs (XRs) can enhance diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate two automated frameworks for binary classification of shoulder XRs.
• To compare the performance and interpretability of an end-to-end deep learning (DL) approach versus a hybrid DL-machine learning (ML) model.

Main Methods:

• A convolutional neural network (CNN) based on VGG19 was trained end-to-end on 4,268 shoulder XRs.
• Hybrid models combined deep features from the CNN with traditional ML classifiers.
• Performance was validated on internal (n=480) and external (n=308) datasets using Receiver Operating Characteristic (ROC) curves and the DeLong test.
• Interpretability was assessed using Grad-CAM and SHAP values.

Main Results:

• Both end-to-end CNN and hybrid CNN-ML models achieved high discriminative performance on internal and external validation sets (AUCs ranging from 0.940 to 0.961).
• No statistically significant differences in performance were found between the two approaches.
• The end-to-end DL model provided more direct visual explainability through saliency maps.

Conclusions:

• Both developed AI frameworks show robustness and high potential for shoulder XR analysis.
• The end-to-end DL approach offers a simpler workflow and enhanced interpretability.
• Further real-world validation and comparison with human readers are required before clinical integration.