An Explainable Artificial Intelligence Framework for Acute Lymphoblastic Leukemia Classification Using Kolmogorov Arnold Networks

Abstract

Acute lymphoblastic leukemia screening from peripheral blood smear microscopy requires accurate and resource efficient models with reliable statistical evaluation. We introduce ResSplit-KAN, a compact interpretable architecture with approximately 1.62 million parameters that combines a residual split multi-scale convolutional backbone with a spline-based Kolmogorov-Arnold Network classification head. Experiments were conducted on a public four-class dataset of 3,242 images using stratified five-fold cross-validation with byte-identical images constrained to the same fold. ResSplit-KAN achieved an accuracy of 99.85 ± 0.10%, balanced accuracy of 99.84 ± 0.09%, macro F1-score of 99.85 ± 0.10%, and expected calibration error of 0.0089 ± 0.0007 after temperature scaling. Fully fine-tuned EfficientNetV2-S and ResNet-50 achieved comparable performance but required substantially larger models and ImageNet pretraining. Ablation experiments quantified the contributions of the residual split backbone, attention modules, KAN head, and preprocessing strategy; a dedicated stain comparison showed that LAB-based segmentation outperformed H&E normalization and individual hematoxylin and eosin channels. Grad-CAM++ explanations were quantitatively assessed using deletion and insertion analysis, while the KAN head provided an exact additive decomposition of class logits. Overall, ResSplit-KAN achieved accurate, well-calibrated, interpretable, and resource efficient classification of individual blood smear cell images without pretraining.

• explainable artificial intelligence
• cross-validation protocol
• leukemia classification
• image segmentation
• Kolmogorov-Arnold network
• lightweight model