VL-OrdinalFormer: Vision–Language-Guided Ordinal Transformers for Interpretable Knee Osteoarthritis Grading

Knee osteoarthritis (KOA) severity assessment using the Kellgren–Lawrence (KL) grading system is essential for clinical decision-making, yet reliable discrimination between adjacent early stages, particularly KL1 and KL2, remains challenging due to subtle radiographic differences and inter-observer variability. This study investigates whether integrating ordinal regression with vision–language semantic alignment can improve fine-grained automated KOA grading. We propose VL-OrdinalFormer, a transformer-based framework that models KL severity as an ordered process and aligns visual features with clinically grounded textual descriptions. The model is evaluated using stratified five-fold cross-validation on the publicly available OAI kneeKL224 dataset (1656 test radiographs). The proposed approach achieves 70.29% accuracy, 70.19% macro F1-score, and 81.61% macro AUROC, outperforming both CNN and standard ViT baselines. Notably, class-wise analysis shows consistent improvements for clinically ambiguous intermediate grades, with gains of +6.6% for KL1 and +19.4% for KL2 compared to the VGG19 baseline. Robustness experiments further demonstrate stable performance under simulated acquisition and projection variability. These results indicate that combining ordinal modeling with vision–language alignment enhances discrimination of subtle disease stages while maintaining interpretability, supporting the potential of the proposed framework for reliable and clinically meaningful KOA grading.