Deep contrastive learning improves identification of early‐stage knee osteoarthritis across multicohort X‐ray datasets

Purpose To develop a Kellgren–Lawrence (K–L) grading recognition framework for knee osteoarthritis (KOA) with enhanced capability for early‐stage detection and to validate its transferability across three independent cohorts. Methods Weight‐bearing anteroposterior knee radiographs were obtained from three datasets: the osteoarthritis initiative (OAI), Wuchuan and Shunyi. The OAI dataset included baseline, 72‐month, and 96‐month follow‐up images, while the Wuchuan and Shunyi datasets were collected from Wuchuan (China) and Shunyi District (Beijing), respectively. Contrastive learning was incorporated into model training to construct the Augmented Dataset‐Wide‐ResMRnet‐Contrastive Loss‐Cross Entropy (AW2C) framework. Results The AW2C framework achieved overall classification accuracies of 83.0%, 82.0% and 80.5% on the OAI, Wuchuan and Shunyi datasets, respectively, with corresponding area under the curve (AUC) of 97.0%, 96.7% and 95.6%. Compared with the baseline model, accuracy for K–L grade 2 improved from 64% to 80%, and discrimination between K–L grades 1 and 2 was notably enhanced. Conclusions The proposed AW2C framework demonstrated robust and transferable performance for automated radiographic K–L grading of KOA, particularly improving recognition of early‐stage and suspected disease. With further optimisation, it holds promise as a reliable tool for large‐scale studies and clinical decision support. Level of Evidence Level III.