Comparative Analysis of CNN and Vision Transformer Architectures for Explainable Brain Tumor Classification Using MRI Images

Abstract

Diagnosis of brain tumor is one of the most critical tasks in medical imaging and early and accurate diagnosis is extremely important for improving the treatment planning and outcome of patients with brain tumors. Brain tumor assessment is achieved with Magnetic Resonance Imaging (MRI), which is commonly used, but manual interpretation of MRI scans can be time-consuming and subject to inter-observer variability. To overcome these difficulties, an explainable deep learning (XDL) framework is proposed that classifies brain tumors into a multiclass classification problem using magnetic resonance imaging (MRI) images. To overcome these challenges, this study proposes an explainable deep learning (XDL) framework that is capable of brain tumor classification as a multiclass classification problem from MRI images. Five state-of-the-art deep learning architectures ResNet50, VGG16, DenseNet121, Vision Transformer (ViT), and a Hybrid CNN ViT model, were tested and compared to determine which one would be the best fit for the classification task. Preprocessing and augmentation of MRI images were performed before training and evaluation the model. Metrics used to evaluate the performance included accuracy, precision, recall, F1 score, the confusion matrix, and Receiver Operating Characteristic Area Under Curve (ROC-AUC). The Hybrid CNN–ViT framework was compared and it was found that the hybrid approach, combining the advantages of CNN for feature extraction and transformer for contextual learning, led to better classification performance. Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-Agnostic Explanations (LIME) were used to explain the best-performing model to improve the transparency and interpretability of the model. To validate that the model attended to clinically relevant tumour regions in classification, the explainability results were obtained. The proposed framework showed high classification accuracy of 98.68% and an average of ROC-AUC score at approximately 0.999, indicating that the framework has good predictive power and high interpretability. The results indicate that the proposed explainable deep learning framework shows great promise of helping clinical diagnosis of brain tumors to be reliable and transparent, by integrating with AI.