Automated Early Detection of Rheumatoid Arthritis in Radiographs Using Neural Networks

TEYI Kodjo Jerôme Sedowo, Sophie Dabo

Abstract: Rheumatoid arthritis (RA) poses diagnostic challenges in resource-limited African settings because early erosive changes on radiographs are subtle and advanced imaging is often unavailable. We propose a joint-centric framework for automated erosion detection on RAM-W600 wrist radiographs, formulated as binary joint-level classification (non-erosive vs. erosive) under severe imbalance (91% non-erosive). An ImageNet-pretrained ResNet-18 trained with focal loss achieves AUC = 0.654, sensitivity = 0.53, and specificity = 0.71 on the test set at threshold 0.314. In addition, joint space width (JSW) analysis validates sub-pixel measurements through controlled synthetic joint space narrowing, revealing strong anatomical variability across joints. Overall, this work provides a reproducible baseline and highlights data and measurement constraints that must be addressed for scalable, trustworthy AI-assisted screening in African healthcare.