Background/Purpose: Rheumatoid arthritis (RA) is a persistent inflammatory disorder inducing joint space narrowing and erosion at both hands and feet. Reading of X-ray images remains extremely time-consuming and subject to interobserver variability.

Methods: Data were obtained from a subset of a large prospective multicentric cohort of RA patients (BCD study). All hand and foot radiographs were scored using the modified van der Heijde score by four trained readers. We developed a deep learning pipeline called RADAR to detect joints and classify radiographs from hands and feet: we employed a YOLO-based object detection for the joints used as input for a classification model. We trained two models: a binary classifier to distinguish erosion and joint space narrowing (present or absence), and a multi-class classifier for joint space narrowing [Mild (scores 1–2) and Moderate (3–4)] and an additional severe category (score 5+) for erosion. The models were implemented in PyTorch. The dataset of 140 patients was split into training, validation, and test sets. We also used the Pytorch Gradcam library to develop explainability heatmaps. Reader reliability was evaluated using both kappa statistics and the intraclass coefficient.

Results: Reliabilities between the four human experts evaluated on 30 radiographs were good (0.7-0.9). A total of 7560 joints were analysed using RADAR . Figure 1 shows examples of joint detection by both readers and RADAR. Table 1 shows that results were excellent for both binary classification and score. This was true in the absence of lesions as well as for elevated scores.

Conclusion: Our results demonstrate the pipeline’s ability to identify the presence and absence of joint damage with high accuracy, giving the opportunity to avoid a large number of human readings in a cohort of patients. The global score is also reliable.

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/radar-a-deep-learning-pipeline-for-automated-scoring-of-joint-damage-in-rheumatoid-arthritis-radiographs/