srivardhan/brain-tumor-detection

Brain Tumor Detection with Grad-CAM Heatmaps

A medical image classification model for detecting brain tumors from MRI scans, built with MobileNetV2 transfer learning and Grad-CAM explainability visualizations.

Results

Metric	Value
Best Validation Accuracy	93.57%
Macro F1-Score	0.94
Architecture	MobileNetV2 (ImageNet V2 pretrained)
Dataset	Hemg/brain-tumor-classification-mri
Classes	glioma, meningioma, no tumor, pituitary
Training Epochs	8

Per-Class Performance

Class	Precision	Recall	F1-Score
Glioma Tumor	0.95	0.89	0.92
Meningioma Tumor	0.90	0.92	0.91
No Tumor	0.92	0.97	0.94
Pituitary Tumor	0.96	0.97	0.97

Training Recipe

Based on published medical imaging research:

• arxiv:2506.09161: MobileNetV2 for brain tumor MRI classification
• arxiv:2307.10506: Two-phase fine-tuning (freeze then unfreeze)
• arxiv:1610.02391: Grad-CAM algorithm for CNN explainability

Key Training Details

• Phase 1 (epochs 1-3): Freeze early layers, train classifier + later conv layers
• Phase 2 (epochs 4-8): Unfreeze all layers, fine-tune with lower LR + cosine annealing
• Adam optimizer, LR=3e-4, weight decay=1e-4
• Data augmentation: flip, rotation, color jitter
• Input: 160x160 RGB, ImageNet normalization

Grad-CAM Heatmaps

Visualizations show which brain regions the model focuses on when making predictions. Target layer: (last convolutional block of MobileNetV2).

Grand Heatmap Grid - All Tumor Types

Shows Original MRI, Grad-CAM activation, and Overlay for 3 samples per class:

Average Activation by Tumor Type

Glioma Tumor Analysis

Meningioma Tumor Analysis

No Tumor (Healthy) Analysis

Pituitary Tumor Analysis

Training Curves

Confusion Matrix

Usage

Files

• best_model.pth - PyTorch model weights
• config.json - Model configuration
• training_history.json - Per-epoch training metrics
• confusion_matrix.png - Test confusion matrix
• training_history.png - Loss/accuracy/LR curves
• heatmaps/ - All Grad-CAM visualizations