Internal Validation of an Uncertainty-Aware Coarse-from-Fine Deep Learning Framework for Knee Osteoarthritis Severity Assessment

Radiographic assessment of knee osteoarthritis remains challenging when adjacent grades on the Kellgren–Lawrence (KL) scale differ only slightly and model predictions are uncertain. We evaluated a “coarse-from-fine” approach based on EfficientNetB3 in a retrospective internal validation study. The hybrid model learned the original KL 0–4 labels and subsequently mapped them to Non-OA, Mild–Moderate OA, and Severe OA. We compared its performance with a model that was trained directly on the same three categories. The dataset comprised 8260 radiographs grouped into 4130 filename-derived grouping identifiers (FDGIs). The paired laterality files were kept within the same subset. The internal evaluation included an FDGI-disjoint held-out test set, grouped cross-validation, repeated training with seeds 42, 123, and 2026, and grouped bootstrap analysis. The mean accuracy was 0.8114 ± 0.0082 for the hybrid approach and 0.8118 ± 0.0027 for direct three-class training with no evidence of general superiority. The entropy-based risk coverage analysis was exploratory in nature and did not establish a validated operational threshold. Grad-CAM provided qualitative visual context, whereas the SHAP results were inconclusive. KneeAI was implemented as a research prototype. Overall, the results support technical feasibility with this dataset, but external and clinical validation is still needed.