Cross-Institutional Five-Class Kellgren-Lawrence Grading of Knee Osteoarthritis via Multitask Deep Learning

Deep learning models for Kellgren-Lawrence (KL) grading often report optimistic performance due to data leakage and fail to generalize across institutions because of domain shift. To address this reproducibility crisis, we introduce KL-FuseNet, a multitask architecture fusing global (ConvNeXt-Base) and local (ResNet-50) features to predict ordinal grades, label distributions, and binary severity (KL≥2). Using strict patient-wise stratified splits on an internal osteoarthritis initiative dataset (n = 8260) and an independent Chinese cohort (n = 2295), we compared zero-shot transfer against selective fine-tuning. KL-FuseNet achieved robust internal agreement (quadratic Cohen's kappa [QWK]: 0.881; accuracy: 70.3%). While external zero-shot deployment revealed a domain gap, with accuracy dropping to 66.1%, our selective fine-tuning protocol significantly bridged this divide, boosting external accuracy to 80.0% and QWK to 0.950, with an AUC of 0.984 for clinically significant osteoarthritis (KL≥2). These results demonstrate that while KL-FuseNet achieves state-of-the-art performance under rigorous evaluation, domain-aware adaptation is essential for clinical utility. This study establishes a reproducible pathway for deploying automated grading models across heterogeneous medical centers.