RadImageNet model classifies scoliosis severity on low-dose biplanar radiographs

A fine-tuned RadImageNet ResNet-50 classified adult scoliosis as none, mild, or severe on AP full-body X-rays with 76.2% accuracy (95% CI 69.2–82.1%) and AUROCs 0.77–0.85. Misclassification mainly occurred between adjacent severity levels; the model does not measure Cobb angle d…

• Retrospective cohort study using 816 adult biplanar radiographs split 80/20 for training/validation and testing; no external validation cohort was available.
• The model achieved its highest F1 score of 0.85 (95% CI 0.79–0.89) for mild scoliosis, with AUROCs of 0.853 (no), 0.766 (mild), and 0.843 (severe).
• The primary limitation is the absence of both external validation and formal interobserver reliability assessment for the final severity categories, limiting generalizability.

RadPigeon summaries are original and for information only. They are not clinical advice.