A hybrid Wavelet-EfficientNetV2-Vision transformer framework for multiclass brain tumor MRI classification

Accurate classification of brain tumors from MRI scans requires models capable of capturing both fine-grained structural details and broader contextual patterns. This study proposes a hybrid Wavelet-EfficientNetV2-Vision Transformer (ViT) framework that integrates multi-resolution frequency enhancement, hierarchical spatial encoding, and global dependency modeling into a unified representation. The Discrete Wavelet Transform (DWT) enriches directional and boundary-aware features, EfficientNetV2 extracts robust localized patterns, and the ViT captures long-range contextual relationships essential for differentiating tumor categories with overlapping visual characteristics. The method was evaluated on the Kaggle-MRI dataset consisting of four classes: glioma, meningioma, pituitary tumor, and no tumor. Experimental results show that the proposed model achieves 98.25% accuracy, with consistent gains in precision, recall, and F1-score compared to standalone CNNs, transformer models, and recent state-of-the-art approaches. Additional analyses, including confusion matrices, ROC-AUC evaluation, and stratified five-fold cross-validation, demonstrate strong class-level stability and generalization. These findings indicate that combining frequency, spatial, and contextual information provides a more discriminative feature space for MRI-based tumor classification. The proposed framework offers a promising direction for developing reliable and high-performance computer-aided diagnosis systems, with future work aimed at multi-center validation and extension to volumetric (3D) modeling.