This project presents an optimized VGG16-based deep learning system for automated brain tumor classification from MRI scans, achieving 94% accuracy and a mean AUC of 0.991 across four classes — Glioma, Meningioma, Pituitary, and No Tumor — using a dataset of 7,023 images. The model incorporates Grad-CAM explainability for clinical transparency, showing radiologists exactly which brain region influenced each decision. A comparative study of five CNN architectures under identical conditions confirmed VGG16 as the best performer, and the complete system was deployed as a clinical web application with tumor segmentation, brain region mapping, and uncertainty detection.

The primary goal of this project was to develop an accurate, explainable, and clinically deployable deep learning system for automated brain tumor classification from MRI images. Specifically, the project aimed to optimize the VGG16 architecture through selective fine-tuning and class-weight balancing to handle imbalanced tumor classes, implement Grad-CAM explainability to make AI decisions transparent for clinical use, and conduct a rigorous comparative evaluation of five CNN architectures under identical training conditions to validate model selection.

The optimized VGG16 model achieved 94% overall accuracy and a mean AUC of 0.991 on 1,600 test images across four classes. Per-class performance showed near-perfect results for Pituitary (F1 = 0.98, AUC = 1.000), No Tumor (F1 = 0.98, AUC = 0.997), and Meningioma (F1 = 0.92, AUC = 0.993), with Glioma as the most challenging class (F1 = 0.88, AUC = 0.975) due to visual overlap with Meningioma. In the five-model comparison under identical conditions, VGG16 ranked first with 89.58% accuracy, followed by MobileNetV2 at 88.89%, InceptionV3 at 84.71%, ResNet50 at 74.47%, and EfficientNetB3 at 42.82%. The Grad-CAM implementation successfully produced accurate tumor localization heatmaps, and the complete system was deployed as a functional web application ready for clinical decision support.