Comparative Evaluation of Single CNN Models, Test-Time Augmentation, and Weighted Ensembling for Knee Osteoarthritis Severity Classification

Knee osteoarthritis (KOA) severity grading via the Kellgren–Lawrence (KL) system suffers from substantial interrater variability, motivating automated deep learning approaches. This study presents a comparative framework combining EfficientNet-B3 and ConvNeXt-Tiny, integrated through test-time augmentation (TTA), prediction averaging over augmented image views, and weighted probability-level ensembling, for four-class KOA severity classification (G0+G1, G2, G3, G4) from knee radiographs. A total of 8260 X-ray images were processed through region-of-interest cropping, contrast-limited adaptive histogram equalisation (CLAHE), and class-aware augmentation, with both learners trained independently using Focal Loss under classweighted and standard sampling. Contrary to the initial hypothesis, ConvNeXt-Tiny alone achieved the highest or nearhighest accuracy (up to 83.8 %), while the weighted ensemble (15 %/85 %) yielded indistinguishable, rather than superior, performance per bootstrap confidence intervals. TTA consistently improved accuracy without requiring retraining or architectural modification. Misclassifications were largely confined to adjacent severity grades, yielding a quadratic weighted kappa of approximately 0.857, indicating preservation of the ordinal KL grading structure – arguably more clinically relevant than raw accuracy. The Minimal (G2) class was the principal source of confusion across confusion-matrix, receiver operating characteristic (ROC), precision–recall (PR), t-distributed stochastic neighbour embedding (t-SNE), and gradient-weighted class activation mapping (Grad-CAM) analyses, reflecting genuine radiographic ambiguity rather than an architecture-specific weakness. Beyond its clinical findings, the proposed dual-branch, multi-resolution inference pipeline and TTA-based probability fusion offer a transferable design pattern for multimedia diagnostic decision-support systems reconciling heterogeneous visual encoders at inference time. These results suggest TTA yields more reliable returns than multi-model ensembling for this task. However, these findings are based on retrospective evaluation using a single OAI-derived dataset, and their clinical applicability remains to be established through external validation and prospective radiologist assessment.