Noise-robust attention-guided deep learning for automated five-class knee osteoarthritis grading from radiographs

Abstract

Knee osteoarthritis (KOA) is commonly assessed using radiographic Kellgren–Lawrence (KL) grading; however, manual evaluation is subjective and prone to inter- and intra-observer variability, particularly in early-stage disease. Although deep learning approaches have shown promise, challenges remain in distinguishing adjacent grades and handling label noise. This study developed and evaluated a noise-robust and attention-guided deep learning framework for automated five-class KOA grading from radiographs. The proposed framework was evaluated on 9786 knee radiographs using stratified cross-validation. It achieved an overall accuracy of 88.52%, a macro-average F1-score of 0.8804, and a macro-average area under the curve of 0.94. Class-wise performance was highest for Grade 0 (F1 = 0.9388) and Grade 4 (F1 = 0.9315), intermediate for Grade 2 (F1 = 0.8718) and Grade 3 (F1 = 0.9106), and lowest for Grade 1 (F1 = 0.7493), with errors concentrated between adjacent early grades. Model interpretability analysis demonstrated that predictions were based on clinically relevant joint regions. However, direct comparison with previous studies is limited due to differences in experimental settings. The proposed framework demonstrates the potential of integrating label-noise mitigation and attention-guided feature learning for automated KOA grading. While the results are promising, further validation on independent multicenter datasets is required to confirm generalizability and support clinical applicability.