Development and validation of an interpretable deep learning model for radiographic grading of knee osteoarthritis severity using X-ray imaging

Zhen DaiChengcheng FengMeng NiZhi TangFangyue ChengJinling Zhang^*

• First Affiliated Hospital of Anhui Medical University, Hefei, China

Accurate radiographic assessment of knee cartilage degeneration remains challenging, particularly for moderate grades and across anatomically heterogeneous subregions. We present X-VIG, an interpretable deep learning framework that integrates paired anteroposterior and lateral knee radiographs through view-specific backbones with deformable convolutions, attention-based feature fusion, and a dual classification–regression optimization tailored to five-grade outputs aligned with clinical grading practice. Additionally, Grad-CAM++ visualizations are incorporated to improve model transparency and support clinical interpretation. In a retrospective cohort (n = 229) with double-blind expert annotations and patient-level splitting (70% training/30% testing), X-VIG achieved consistently strong discrimination across five anatomical subregions (AUCs: 0.9646–0.9945), with the medial femoral condyle reaching AUC 0.9945 and F1 0.9663. Multimodal fusion of anteroposterior and lateral views provided a mean AUC gain of 5.6 percentage points over single-view models, underscoring complementary value across projections. In a 100-case reader comparison, the model outperformed junior radiologists on all metrics (average precision +37.8 Zhen Dai et al. Interpretable Knee Cartilage Grading and recall +35.4 percentage points), matched or surpassed senior performance in most subregions, and processed paired views in 27 seconds per case—an 82% reduction in reading time—suggesting meaningful workflow benefits. These results indicate that X-VIG delivers accurate, efficient, and interpretable grading on routine X-rays, positioning it as a credible second reader to standardize reporting and reduce variability in clinical practice. The implementation code is publicly available at: https://github.com/AngelaK-code/KneeCartGRAD.