Evaluation of an Explainable Convolutional Neural Network with Grad-CAM for Brain Tumor Classification from Magnetic Resonance Imaging

Abstract

Correct classification of the tumor is crucial for the diagnosis and clinical management of neuro-oncology. Early tumor typing can help the clinician prioritize follow-up testing and plan for appropriate care. This study introduces an explainable brain tumor classification framework that combines a convolutional neural network (CNN) and Gradient-weighted Class Activation Mapping (Grad-CAM) for not only predictive accuracy but also visual interpretability derived from magnetic resonance imaging ( MRI). The MRI dataset consists of 7023 images that belong to four classes: glioma, meningioma, pituitary tumor, and no tumor, and were preprocessed, resized, normalized, and then split into training, validation, and testing sets. The suggested CNN-Grad-CAM was tested with three popular deep learning architectures: VGG16, ResNet101, and ShuffleNetV2. The results showed that CNN-Grad-CAM achieved the highest accuracy (94.75%), precision (94.99%), recall (94.75%), and F1-score (94.83%). The best baseline, ResNet101, was the one that had the highest accuracy of 93.26% and F1 score of 93.21%. Grad-CAM was also used to create heat maps that illustrate which areas of the image are important for the model's output, thus forming an interpretable foundation for analyzing the model's classification decisions. The visualizations obtained, together with quantitative classification results, focused on the tumor-relevant regions in representative MRI scans. Overall, the results show that the proposed CNN-Grad-CAM framework is a robust, explainable computer-aided method for brain tumor classification using MRI images with four classes.