Overview
Brain tumors are life-threatening neurological conditions requiring prompt and accurate diagnosis. MRI provides detailed structural insight but manual interpretation is time-consuming and error-prone. This project automates four-class tumor classification using state-of-the-art CNN architectures with Grad-CAM visualization for clinical interpretability.
The four classes: Glioma (aggressive cancerous), Meningioma (non-cancerous, slow-growing), Pituitary Tumor (hormone-affecting), No Tumor.
Dataset
Combined Figshare (3,064 images — Glioma, Meningioma, Pituitary) and BR35H datasets (No Tumor class). Images normalized to [0,1], resized to 128×128, batched at 32.
Architectures Compared
VGG16
Simple sequential structure. Best performing overall — simple yet highly effective.
ResNet50
Solves vanishing gradient via residual blocks. Highly robust performance.
VGG19
Slightly deeper than VGG16. Balanced performance.
DenseNet121
Feature reuse across layers. Compact and efficient. Sharpest Grad-CAM.
InceptionV3
Parallel convolutions capturing multi-scale MRI features.
EfficientNetB0
Best efficiency-accuracy tradeoff for resource-constrained environments.
ResNet152
Deeper variant. Slightly lower due to overfitting on dataset size.
Custom CNN
PyTorch baseline: 3 conv layers (32→64→128 filters) + 2 FC layers.
Quantitative Results
All pretrained models achieved ROC-AUC 0.99–1.00
Grad-CAM Visualization
Grad-CAM applied to all architectures to visualize which MRI regions drive each model's prediction. Critical for clinical trust.
Real Grad-CAM outputs: Red/orange = high activation · Blue = low activation · Smaller hotspot = more precise localization
Conclusion
All pretrained CNN architectures significantly outperformed the custom baseline. VGG16 leads on accuracy (98%) while DenseNet121 produces the most clinically useful Grad-CAM visualizations. Future work: ensemble approaches combining VGG16 + DenseNet121, 3D volumetric analysis using full MRI scan sequences, and clinical validation with radiologist feedback.