Rheumatoid arthritis (RA) is a chronic autoimmune disease leading to joint damage, which is typically quantified using Sharp/van der Heijde (SvH) scores on hand and foot X-rays. Deep learning–based artificial intelligence (AI) enables automated SvH scoring through (1) global models that process whole radiographs and (2) per-joint pipelines that first localize joints and then score erosions and joint space narrowing with greater interpretability. Remaining challenges include limited expert-labeled data, severe class imbalance, and rater variability, which restrict the robustness and generalizability of these algorithms. Emerging strategies—such as semi-/self-supervised pseudo-labeling, advanced segmentation architectures, and symmetry-aware modeling—combined with AI deployed as a calibrated “second reader,” may improve sensitivity to subtle progression while maintaining essential human oversight. Harnessing AI could deliver more accurate and efficient radiographic assessment, thereby supporting better research as well as clinical care for patients with RA.
Meng et al. (Thu,) studied this question.