Standard and decoupled knowledge distillation for lightweight radiographic knee osteoarthritis grading and triage

Abstract This retrospective study compares standard knowledge distillation (KD) and decoupled knowledge distillation (DKD) fo​r lightweight artificial intelligence-based​ radiographic knee osteoarthritis grading. GhostNetV2 students learned five Kellgren–Lawrence grades from a Swin-Tiny teacher and were evaluated against a cross-entropy baseline using an Osteoarthritis Initiative-derived dataset of 8260 images from 4130 patients with patient-disjoint partitions. Across three​ runs, mean binary triage accuracy was 84.04%, 84.82%, and 84.82% for baseline, standard KD, and DKD. Thresholds targeting at least 90% validation specificity yielded test sensitivities of 72.77%, 76.56%, and 75.50%, with specificities of 92.09%, 90.94%, and 91.12%; these operating points are not specificity-matched. Mean five-class accuracy was 64.63%, 67.63%, and 67.71%. Separate illustrative checkpoints showed reduced KL-1-to-KL-2 overcalls with DKD (28.0 to 18.9%) but lower KL-0 recall (90.6 to 78.7%). Calibration, exploratory decision curves, and patient-cluster uncertainty intervals were evaluated. Each student contains 4.88 million parameters. Distillation improved mean performance internally, but decoupling did not demonstrate an additional sensitivity benefit over standard KD. No external cohort or clinician-supported decision policy was evaluated; clinical utility and population-level stability remain unestablished.