Taylor-series expanded Kolmogorov–Arnold Network for medical imaging classification

Effective and interpretable classification of medical images remains a key challenge in computer-aided diagnosis, particularly in data-scarce and resource-constrained clinical settings. This study introduces spline-based Kolmogorov–Arnold Networks (KANs) for accurate classification of limited and heterogeneous medical imaging datasets. Three variants were developed: SBTAYLOR-KAN, integrating B-splines with Taylor series expansion; SBRBF-KAN, combining B-splines with radial basis functions; and SBWAVELET-KAN, embedding B-splines within Morlet wavelet transforms. The models were evaluated on brain MRI, chest X-ray, and tuberculosis chest X-ray datasets without any image preprocessing beyond the datasets’ original published forms, demonstrating the ability to learn directly from unprocessed raw data. Cross-dataset validation, robustness checks, and data reduction analysis confirmed strong generalization and stability. Among all variants, SBTAYLOR-KAN demonstrated superior performance, achieving 93.04% accuracy (F1: 93.02%) on brain tumor, 96.37% accuracy (F1: 97.00%) on COVID-19 chest X-ray, and 98.93% accuracy (F1: 96.48%) on TB chest X-ray datasets. Trained on only 30% of available data, it retained over 86% accuracy on brain MRI and over 92% on chest X-rays. With only 2 , 872 trainable parameters, SBTAYLOR-KAN achieves competitive or superior performance against substantially larger CNNs, while recording the fastest training (13.71, 12.35, and 21.34 min across datasets) and lowest inference latency (2.69–2.75 ms per image), making it 2– 3 × faster to train and up to 1 . 8 × faster at inference than the other variants. This efficiency reduces computational cost and hardware dependence, making the framework ideal for portable, real-world diagnostic systems. Grad-CAM enhances interpretability for low-resource clinical deployment. The source code is publicly available at Fatema (2026).