Early-grade collapse in deep knee osteoarthritis grading: imbalance-aware training under controlled ordinal regularization

Automated radiographic grading of knee osteoarthritis is summarized by accuracy or quadratic weighted kappa (QWK), which can hide failure at Kellgren–Lawrence (KL) Grade 1. We studied this early-grade collapse using ImageNet-pretrained EfficientNet-B4 models (380 pixels) under a fixed protocol on the official Knee Osteoarthritis Dataset with Severity split. File names follow a seven-digit-identifier-plus-side pattern consistent with participant identifiers; interpreted this way, the split is subject-disjoint (4130 identifiers, both knees in the same split). A five-seed 2 × 2 factorial design crossed weighted cross-entropy with mild class-balanced sampling; a five-seed paired comparison evaluated CORAL-style ordinal regularization with pairwise ranking. QWK, the primary endpoint, did not differ detectably between configurations (0.824–0.838). Plain cross-entropy recalled only 0.118 ± 0.068 of Grade-1 knees. Class weighting and sampling raised Grade-1 recall (main effects +0.091 and +0.083; Holm-adjusted p = 0.016 and 0.029), reducing Grade-1 specificity (−0.034 and −0.029). Threshold-free Grade-1 discrimination was unchanged (average precision 0.350–0.360; AUROC 0.755–0.763), and a validation-selected Grade-1 decision offset applied to plain cross-entropy achieved comparable precision–recall operating points. Ordinal regularization improved no endpoint over the imbalance-aware baseline (QWK difference −0.005, 95% CI −0.019 to 0.008); ordinal weight had no detectable effect across three seeds, while blending ordinal-head probabilities monotonically lowered Grade-1 recall. Validation-tuned class-bias correction improved validation accuracy in all six runs but reduced test accuracy in four, left it unchanged in one, and improved it in one (mean test change −0.006). Retraining after removing perceptual-hash near-duplicate candidates did not change performance. Within this benchmark, early-grade collapse is largely a decision-threshold phenomenon: imbalance handling shifts the Grade-0/Grade-1 operating point without improving Grade-1 separability. Grade-1 sensitivity and specificity should accompany aggregate agreement.