A Deep Learning Approach Based on XAI and ViT-GRU Hybrid Model for Brain Tumor Classification Using MRI Images

Brain tumors (BT), resulting from uncontrolled cell proliferation in the central nervous system, present significant diagnostic challenges due to their anatomical complexity and interpatient variability. Early and accurate classification using MRI scans is crucial for effective clinical intervention. This study proposes a hybrid Vision Transformer-Gated Recurrent Unit (ViTGRU) model enhanced with Explainable Artificial Intelligence (XAI) to improve classification accuracy and interpretability. The ViT backbone captures spatial dependencies through selfattention mechanisms, while the GRU layer models temporal patterns in sequential features. Two publicly available MRI datasets were utilized: a binary classification dataset (253 images) and a multi-class dataset (3,264 images). Pre-processing included grayscale-to-RGB conversion, normalization, and data augmentation. The model was trained using the AdamW optimizer and evaluated against pretrained CNNs (ResNet50, InceptionV3) and ViT variants (ViT-S/16, ViT-L/16). Experimental results demonstrated state-of-the-art performance, achieving 92% accuracy (binary) and 87.73% accuracy (multi-class), outperforming baseline models. Integration of Local Interpretable Modelagnostic Explanations (LIME) generated heatmaps highlighting tumor regions, enhancing clinical interpretability. The proposed framework combines robust classification with transparency, offering a reliable tool for BT diagnosis. Future work will focus on expanding datasets and deploying the model in clinical workflows.