Explainable Deep Learning Framework for Multi-Class Brain Tumor Classification Using VGG16 and Grad-CAM Visualization

This study proposes an explainable deep learning framework for the automated detection and multi-class classification of brain tumors from MRI images, addressing key challenges in diagnostic accuracy, generalizability, and clinical interpretability. The framework employs a transfer learning–based Convolutional Neural Network (CNN) using the VGG16 architecture, fine-tuned on a balanced dataset comprising 6,484 MRI images collected from three publicly available repositories Figshare, SARTAJ, and Br35H. The dataset includes four classes: glioma, meningioma, pituitary, and no tumor, with equal class representation to ensure unbiased learning. Preprocessing was performed using the Python Imaging Library (PIL) to resize, normalize, and enhance image quality, while Kera's-based data augmentation introduced random variations in brightness and contrast to improve robustness against overfitting. The fine-tuned VGG16 model, with frozen early convolutional layers and retrained dense layers, achieved an overall classification accuracy of 95%, outperforming comparable deep architectures such as ResNet50 (93.5%), DenseNet121 (94%), and InceptionV3 (93%).Comprehensive performance evaluation through precision, recall, F1-score, and ROC–AUC analysis confirmed consistent multi-class discrimination, achieving a macro-average AUC of 0.97. Furthermore, Grad-CAM visualizations provided clear, class- specific heatmaps highlighting tumor-affected regions, thereby enhancing model transparency and diagnostic reliability. The integration of quantitative performance metrics with visual interpretability demonstrates that the proposed VGG16- based framework delivers clinically explainable, efficient, and accurate diagnostic support, reducing inter-observer variability and assisting radiologists in early brain tumor detection and treatment planning.