Evaluating Kolmogorov-Arnold Network Implementations for Medical Ultrasound Image Classification: A Trade-off Analysis of Computational Cost, Interpretability, and Diagnostic Safety.

Abstract

To compare pykan and Lightweight KAN with MLP and Linear classifier heads for focal liver lesion classification on B-mode ultrasound, with emphasis on diagnostic sensitivity, interpretability and computational cost. A frozen ResNet18 feature extractor was combined with 4 classifier heads and evaluated on 735 annotated liver ultrasound images (Benign, Malignant, Normal) using 5-fold stratified cross-validation. Weighted cross-entropy was applied to address class imbalance. Performance was assessed using Accuracy, F1-Macro, F1-Weighted, AUROC, class-wise recall, malignant-to-normal misclassification count, inference latency, and pykan-specific spline activeness analysis. Overall metrics were similar across models, and Wilcoxon Signed-Rank testing showed no statistically significant pairwise differences (p > .05). However, clinically relevant class-level differences emerged. pykan achieved the highest Benign recall (65.5%) and the lowest malignant-to-normal error rate (5/435, 1.1%). Lightweight KAN achieved the highest Malignant recall (82.1%) and markedly faster inference (0.69 ms/image) than pykan (4183.67 ms/image). Aggregate metrics alone do not capture clinically meaningful trade-offs among KAN implementations. Lightweight KAN is attractive for rapid screening, whereas pykan shows a directional trend toward fewer malignant-to-normal errors and a more interpretable profile, though this trend did not reach statistical significance and requires confirmation in larger studies.