Development and Validation of an Interpretable Deep Learning Model for Automated Kellgren-Lawrence Grading of Knee Osteoarthritis Using X-Ray Imaging.

Abstract

Accurate and reproducible Kellgren-Lawrence grading of radiographic knee osteoarthritis remains challenging, particularly for intermediate grades and anatomically heterogeneous compartments. We present X-VIG, an interpretable deep learning framework integrating paired anteroposterior and lateral knee radiographs via view-specific backbones with deformable convolutions, attention-based feature fusion, and dual classification-regression optimization aligned with five-grade clinical practice. Grad-CAM++ visualizations enhance model transparency and support clinical interpretation. In a retrospective cohort (n = 229) with consensus reference-standard KL labels established by double-blinded expert review and patient-level splitting (70%/30%), X-VIG achieved AUCs of 0.9646-0.9945 across five anatomical subregions. Multimodal fusion yielded a mean AUC gain of 5.6 percentage points over single-view models. In a reader study of 50 held-out patients comprising 100 individual radiographs, X-VIG outperformed junior radiologists, matched or surpassed senior performance in most subregions, and reduced reading time by 82% (27 s per case). These results position X-VIG as a credible second reader for standardizing the reporting of knee osteoarthritis in routine clinical workflows.