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

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Rheumatology

Background:

• Rheumatoid arthritis (RA) diagnosis relies on radiographic assessment, often subjective.
• Automated analysis of hand radiographs can improve diagnostic accuracy and efficiency.
• Deep learning offers potential for objective and rapid RA classification.

Purpose of the Study:

• To develop an interpretable deep learning framework for automated rheumatoid arthritis (RA) classification in hand radiographs.
• To elucidate the decision-making patterns of the deep learning model.
• To enable clinical translation through a web-based deployment.

Main Methods:

• A retrospective study analyzed 1,655 hand radiographs from RA patients and healthy controls.
• A Visual Geometry Group (VGG)-8 convolutional neural network was trained on enhanced X-ray images.
• Interpretability was assessed using Gradient-weighted Class Activation Mapping (Grad-CAM) and Shapley Additive Explanations (SHAP).
• A web application was developed using Streamlit for clinical practicality.

Main Results:

• The classifier achieved high training performance (AUC=0.99, accuracy=0.94) and generalizable testing metrics (AUC=0.81, accuracy=0.74).
• Interpretability analysis highlighted joint space stenosis, bone erosion, and trabecular changes as key pathological indicators.
• Metacarpophalangeal and wrist joints were identified as critical predictive features by SHAP analysis.
• The web application demonstrated utility in assisting RA diagnosis from hand X-rays.

Conclusions:

• The developed framework integrates deep learning with interpretable insights for RA diagnosis in hand radiographs.
• This approach aids in the early detection of RA and enhances clinician proficiency.
• The model provides diagnostic assistance and educational insights into RA radiographic markers.