Abstract
BACKGROUND: Parkinson's Disease diagnosis remains challenging due to subtle early brain changes. Deep learning approaches using brain scans may assist diagnosis, but optimal architectures remain unclear. This study applies Convolutional Kolmogorov-Arnold Networks (ConvKANs), which use flexible mathematical functions for feature extraction, to classify Parkinson's Disease from structural brain scans. METHODS: We implemented the first three-dimensional ConvKAN architecture for medical imaging and compared performance against established deep learning models, including Convolutional Neural Networks, Vision Transformers, and Graph Convolutional Networks. Three publicly available datasets containing brain scans from 142 participants (75 with Parkinson's Disease, 67 healthy controls) were analyzed. Models were evaluated using both two-dimensional brain slices and complete three-dimensional volumes, with performance assessed through cross-validation and independent dataset testing. RESULTS: Here we show that two‑dimensional ConvKAN achieved an AUC of 0.973 for Parkinson's‑disease detection, outperforming a pretrained ResNet (AUC 0.878, p = 0.047). On the early‑stage PPMI hold‑out set, the three‑dimensional variant generalised better than the two‑dimensional model (AUC 0.600 vs 0.378, p = 0.013). Furthermore, ConvKAN required 97% less training time than conventional CNNs while maintaining superior accuracy. CONCLUSIONS: ConvKAN architectures offer promising improvements for Parkinson's Disease detection from brain scans, particularly for early-stage cases where diagnosis is most challenging. The computational efficiency and strong performance across diverse datasets suggest potential for clinical implementation. These findings establish a framework for artificial intelligence-assisted diagnosis that could support earlier detection and intervention in Parkinson's Disease.