FPGA-Based Implementation Using Graph Spectral Features for Multiclass Brain Tumor Classification

The growth of brain tumors (BTs) is often highly aggressive, resulting in low survival rates. Therefore, early and accurate detection is critical for improving patient outcomes. Classifying tumors from magnetic resonance imaging (MRI) scans is challenging due to high interclass similarity and the complex, non-Euclidean nature of tumor structures. Conventional techniques may struggle to capture both spatial relationships and neighborhood information, limiting accuracy. To address these challenges and capture a holistic view of the brain, a graph signal-processing (GSP)-based feature extraction approach utilizing the spectral graph wavelet transform (SGWT) is proposed. This method effectively retains both spatial and spectral features, enhancing classification performance. Unlike traditional techniques such as the gray level co-occurrence matrix (GLCM), local binary patterns (LBPs), and discrete wavelet transform (DWT), which primarily focus on neighborhood information or multiscale features, SGWT leverages graph-based signal processing to integrate spatial and spectral relationships, leading to superior classification accuracy. The proposed method is implemented on PYNQ-ZU hardware to assess its hardware compatibility. Experimental results on two multiclass datasets—one with three tumor classes and another with fifteen—show that the light gradient boosting machine (LGBM) classifier achieves 99.97% accuracy on the three-class dataset, while the categorical boosting (CatBoost) classifier attains 99.98% accuracy on the fifteen-class dataset. Fivefold cross-validation and linear discriminant analysis (LDA) are applied to prevent overfitting. The proposed method outperforms state-of-the-art techniques in terms of accuracy and efficiency. Furthermore, its lightweight nature makes it highly suitable for real-time BT classification.