Brain Cancer MRI Classification

A deep learning project for classifying brain MRI scans into three categories: glioma, meningioma, and other brain tumors using Convolutional Neural Networks (CNN) with PyTorch.

Project Overview

This project implements a custom CNN architecture to classify brain MRI images into three distinct categories:

• Brain Glioma - A type of tumor that occurs in the brain and spinal cord
• Brain Meningioma - A tumor that arises from the meninges
• Brain Tumor - Other types of brain tumors

The model achieves 86.9% accuracy on the test set with strong performance across all three classes.

Key Results

Activation Function Comparison

Activation	Final Epoch	Train Loss	Val Loss	Train Acc (%)	Val Acc (%)
LeakyReLU	50	0.450	0.356	81.67	86.47
SiLU	50	0.477	0.375	80.37	85.48
ReLU	50	0.459	0.416	81.48	84.71
ELU	50	0.502	0.749	79.38	63.70

The LeakyReLU activation function achieved the best validation accuracy of 86.47%, closely followed by SiLU and ReLU. ELU showed signs of overfitting with significantly lower validation accuracy.

Model Architecture

Custom CNN architecture (MRIClassifier) with the following layers:

1. Convolutional Block 1: Conv2d (3→32) → ReLU → MaxPool
2. Convolutional Block 2: Conv2d (32→64) → ReLU → MaxPool
3. Convolutional Block 3: Conv2d (64→128) → ReLU → MaxPool
4. Flatten Layer
5. Fully Connected Block 1: Linear → ReLU → Dropout(0.5)
6. Fully Connected Block 2: Linear → ReLU → Dropout(0.5)
7. Output Layer: Linear (→3 classes)

Input: 224×224×3 RGB images (grayscale converted to 3 channels)
Output: 3-class softmax predictions

Technologies & Dependencies

• Python 3.x
• PyTorch - Deep learning framework
• torchvision - Image transformations and datasets
• NumPy - Numerical computations
• Pandas - Data manipulation and analysis
• Matplotlib - Visualization
• PIL (Pillow) - Image processing
• Kagglehub - Dataset management

Dataset

• Total Images: 6,056 MRI scans
• Classes: 3 (brain_glioma, brain_menin, brain_tumor)
• Train/Val/Test Split: Stratified split for balanced evaluation
• Test Set Size: 909 images

Data Preprocessing

transforms = v2.Compose([
    v2.Resize((224, 224), antialias=True),
    v2.Grayscale(num_output_channels=3),
    v2.ToImage(),
    v2.ToDtype(torch.float32, scale=True),
    v2.Normalize(mean=[0.485, 0.456, 0.406], 
                 std=[0.229, 0.224, 0.225])
])

Activation Function Comparison

The project includes experiments with different activation functions:

Activation	Final Val Accuracy	Model File
SiLU	85.48%	ClassyModel_SiLU.pth
LeakyReLU	~86%	ClassyModel_LeakyReLU.pth
ELU	63.70%	ClassyModel_ELU.pth

See BrainCancerMRI_diffacfunc.ipynb for detailed comparisons.

Training Details

• Loss Function: CrossEntropyLoss
• Optimizer: Adam
• Epochs: Multiple training runs with different configurations
• Dropout: 0.5 (applied in fully connected layers)
• Device: CPU/CUDA (automatic detection)

Contributing

This project was developed as part of a Data Mining and Machine Learning coursework at Heriot-Watt University.