Explainable brain tumor diagnosis with SHAP-enhanced graph wavelet features on FPGA

Brain tumors are among the most life-threatening neurological disorders, where timely and accurate diagnosis is critical for effective treatment planning and improved patient survival. Despite substantial advances in automated brain tumor analysis, reliable classification from magnetic resonance (MR) images remains challenging due to complex spatial dependencies, highly irregular tumor structures, and non-Euclidean relationships that conventional models inadequately capture. To address these limitations, this study proposes a novel graph signal processing (GSP)–based framework that models brain MR images as graph signals, enabling joint representation of spatial relationships, neighborhood connectivity, and spectral characteristics. Discriminative features are extracted using the spectral graph wavelet transform (SGWT) with Abspline and Mexican Hat wavelet kernels to capture multi-scale spectral information. The extracted graph-domain features are classified using Light Gradient Boosting Machine and Categorical Boosting classifiers, achieving robust performance with low computational complexity. Experimental evaluations on the Br35H and Kaggle-4600 brain MRI datasets demonstrate high classification accuracies of 97.33% and 98.11%, respectively. To enhance transparency and clinical trust, SHAP-based explainable artificial intelligence (XAI) is incorporated to identify the most influential graph-derived features. Furthermore, the proposed framework is implemented on an Xilinx PYNQ-ZU FPGA platform, achieving real-time inference with a latency of 0.0144 s and a compact model size of only 0.0031 million parameters. These results demonstrate that the proposed approach is accurate, explainable, and hardware-efficient, making it suitable for resource-constrained and IoT-enabled healthcare environments. • Propose a novel GSP-based framework for brain tumor detection. • Introduce weighted graph models to capture neighborhood spatial connectivity. • Apply Abspline and Mexican Hat wavelets as graph-based measurement operators. • Deploy real-time system on Xilinx PYNQ-ZU with 0.0031M parameters. • Reach 98.11% accuracy with SHAP-based interpretability on MRI tumor datasets.