Deep Learning Classification of Rheumatoid Arthritis in Hand Radiographs Interpretability Insights and Web Application.

Abstract

Purpose: To establish an interpretable deep learning framework for automated classification of rheumatoid arthritis (RA) in hand radiographs, with emphasis on elucidating model decision-making patterns and enabling clinical translation through web-based deployment.

Patients and methods: A retrospective multicenter study analyzed 1,655 hand radiographs (809 RA patients, including early RA cases, and 846 healthy controls). Enhanced data (random rotation, brightness/contrast adjustment) was applied to the collected X-ray images to improve the model's generalization ability and performance. Subsequently, A lightweight Visual Geometry Group (VGG)-8 convolutional neural network was trained and validated using processed hand X-ray images. This model has the ability to distinguish RA patients from healthy controls. The interpretability of the model was systematically evaluated using both Gradient-weighted Class Activation Mapping (Grad-CAM) and Shapley Additive Explanations (SHAP). Finally, a web application was developed using Streamlit that supports JPEG input, helps to address the clinical practicality of the model.

Results: For distinguishing RA patients from healthy individuals, the classifier achieved excellent training performance (AUC=0.99, accuracy=0.94) and generalizable testing metrics (AUC=0.81, accuracy=0.74). Specifically, the model was successfully constructed and demonstrated good performance in external validation. Interpretability analysis revealed areas of pathological significance, with Grad CAM heatmaps highlighting structural abnormalities (joint space stenosis, bone erosion, trabecular structural changes), and SHAP values analysis identifying metacarpophalangeal and wrist joints as key predictive features. A web application developed using Python and Streamlit framework can assist in the diagnosis of RA hand X-ray images in clinical practice.

Conclusion: This work advances clinical diagnosis, including early RA patients, by integrating deep learning with interpretable decision paths in hand radiographic analysis, while helping clinicians to use the model more proficiently. The framework provides both diagnostic assistance and educational insights into RA radiographic markers.