KneeAI Reproducibility Package for Uncertainty-Aware Knee Osteoarthritis Severity Assessment

The Keras EfficientNet implementation includes input rescaling within the model; its `preprocess_input` function does not perform additional ImageNet mean–standard deviation normalization. 4. Apply training-only data augmentation using the documented rotation, width and height shifts, zoom, horizontal flipping, and nearest-neighbor filling. Do not augment the tuning or held-out internal test images. 5.

Initialize the ImageNet-pretrained EfficientNetB3 backbone and construct the corresponding classification head. For the hybrid formulation, train with the original five KL grades (KL 0–4). For the direct comparator, train using the predefined three-category severity mapping. 6.

Follow the documented staged transfer-learning protocol: an initial warm-up with the backbone frozen, followed by fine-tuning of the designated upper backbone layers. Use the archived training configuration and checkpoint-selection procedure rather than substituting a new training schedule. 7. For the historical five-class optimization, use the documented Optuna Tree-structured Parzen Estimator search and its recorded hyperparameter configuration.

For the principal matched comparison, use the fixed, shared experimental configuration and the three documented random seeds (42, 123, and 2026). Select checkpoints using tuning-subset performance without using the held-out internal test subset for model selection. 8. Evaluate the original five-class KL formulation and the common three-class severity output space.

For the hybrid model, map the five-class categorical prediction into Non-OA (KL 0–1), Mild–Moderate OA (KL 2–3), or Severe OA (KL 4). Aggregate the corresponding five-class probabilities for probability-based three-class analyses. Compare the hybrid formulation with the matched directly trained three-class model. 9.

Perform the documented internal-validation analyses, including five-fold stratified group cross-validation within the development cohort, the FDGI-disjoint held-out internal test, matched three-seed comparison, and FDGI-level grouped bootstrap analysis. Report the documented classification, discrimination, calibration, and uncertainty metrics, keeping historical fixed-artifact analyses separate from the principal matched experiment.experiment.