An Advanced XAI Framework for MRI-Based Glioma Classification Using Vision Transformer and CNN-Based Approaches

Abstract

This paper presents an advanced eXplainable Artificial Intelligence (XAI) framework for Magnetic Resonance Imaging (MRI)-based glioma classification, designed to bridge the gap between high-performing Deep Learning (DL) models and interpretability in clinical diagnostics. The proposed framework leverages three Convolutional Neural Network (CNN) models (ResNet50V2, EfficientNet-B6, and Xception) and a Vision Transformer (ViT) model to differentiate between Low-Grade Gliomas (LGG) and High-Grade Gliomas (HGG). A Max-Voting strategy integrates the predictions of these models to enhance classification reliability. To promote transparency, the framework incorporates four XAI methods (LIME, SHAP, Integrated Gradients, and GradCAM), providing interpretive insights by highlighting salient features and regions within the MRI scans. Experiments using the BraTS2019 dataset demonstrate that the suggested framework achieves a high classification accuracy of 98.89%, while also offering visual attention maps and interpretative information flow visualizations that support radiologists in clinical decision-making. This work not only advances accurate glioma classification but also emphasizes the role of transparency in AI, highlighting how explainable AI has the potential to revolutionize medical imaging.