MRI Alzheimers Classification using Deep Learning

Deep learning model for classifying brain MRI scans into dementia severity categories using a modified ResNet-18 convolutional neural network.

Dataset

The dataset contains labeled MRI scans categorized into four classes:

Class Description

  |     Non Demented     |     Very Mild Demented   |     Mild Demented     |     Moderate Demented 

Dataset split:

Split	Images
Training	23,788
Validation	5,097
Test	5,097

Images are organized into folders and loaded using torchvision.datasets.ImageFolder.

Model Architecture

The project uses a modified ResNet-18 architecture for its residual blocks, solving the vanishing gradient problem.

Modifications:

• MRI scans are grayscale, so the first convolution layer was modified
• Per-Image Z-Score Normalization
• Removes brightness variation between scans.

x_norm = (x - μ) / σ

Activation: ReLU

Loss Function: Weighted Cross Entropy Loss is used to address class imbalance.

L = - Σ wi * yi * log(ŷi)

Optimizer: AdamW

Tools

• Python
• PyTorch
• Torchvision
• CUDA
• Optuna(Baysian)
• NumPy