Evidence-Centered Radiographic Grading of Knee Osteoarthritis: A Review of Deep Learning, Ordinal Modeling, Fusion, and Explainable AI

Abstract

Automated severity grading of knee osteoarthritis (KOA) has progressed from handcrafted descriptors to deep networks, ordinal objective, attention mechanism and multimodal learning, and explainable artificial intelligence. The source presentation motivating this review focuses on fusion of complementary structur-al and textural radiographic evidence and hybrid learning and interpretable predic-tion. It is well supported recently in literature but some persistent problems can be noted: the classes are not equally sized, the separation of adjacent Kellgren–Lawrence (KL) grades are challenging, performance of the model can be below external validation, and the quality of the explanations is not measured quantita-tively. In this review, we categorize the methods used for grading KOA into handcrafted/radiomics, convolutional and transfer-learning systems, ordinal or hi-erarchical models, transformer and vision–language and fusion-oriented pipe-lines. According to comparative analysis, the contributions of architecture alone are hard to interpret, as there is a strong influence from composition of datasets, labeling practice, projection, preprocessing and validation design on reported measures. Evidence centered fusion is thus not a single network architecture, but a mixture of complementary representations, ordinally aware predictions, calibra-tion, and clinically grounded XAI. Priority directions for clinically credible de-ployment identified in the review include clinically relevant measures of uncer-tainty, quantitative assessment of the quality of explanations, imaging integrated with clinical or longitudinal evidence, leakage-safe splitting, and multi-center evaluation.