Abstract / Summary
Objectives: To develop a model for the classification of knee OsteoArthritis (OA) on radiographs, which is computationally efficient and severity ordered. Method: This study focuses on class imbalance, adjacent-grade confusion and high computational cost of multi-model hybrid systems. The proposed SOBR-KneeNet (Severity Ordered Boundary Residual Learning) includes a pretrained ConvNeXt-Tiny encoder and three heads for supervised learning: classification, cumulative-ordinal and residual-estimation. Severity structure in training is provided by auxiliary ordinal and residual objectives and the final inference distribution through validation calibration. In experiments, 8,260 Osteoarthritis Initiative knee radiographs are used, where the 5,778, 826 and 1,656 are used as training, validation and test partitions, respectively. The training protocol consists of the AdamW optimization, progressive backbone unfreezing, mixed precision, cosine learning-rate decay and horizontal-flip test-time augmentation. Findings: The test accuracy of SOBR-KneeNet was 68.90%, balanced accuracy was 71.20%, macro F1-score was 70.38%, weighted F1-score was 68.86%, grade mean absolute error was 0.3904, quadratic weighted kappa was 0.8450 and within one grade accuracy was 95.59%. The accuracy was 2.11 % higher than existing study, 2.59% higher than DenseNet-121 and 3.54 % higher than ResNet-50 V2. The percentage with grade-4 recall was 90.20% and all errors were within two grades of the correct level. Novelty: Severity information is fed into optimization directly, without the need for a transformer branch, multi-model ensemble or iterative post-processing, through fixed cumulative boundaries. Validation-selected inference retained the calibrated 5-class head, with the ordinal and residual branches serving as low-cost training regularizers. The design offers one-backbone solution that has 28.63 million parameters and a fast epoch execution time. Keywords: knee osteoarthritis, Kellgren-Lawrence grading, ordinal learning, ConvNeXt, CLAHE, multi-task learning