Purpose

This pull request introduces a Vision Transformer (ViT) model designed for the classification of Alzheimer’s Disease using MRI scans from the ADNI dataset. The model leverages the self-attention mechanism inherent to Transformers to extract meaningful features across different regions of the brain, aiming to differentiate between 'Normal' and 'Alzheimer’s Disease' categories.

Description

Task Chosen: Implement a ViT-based classifier for Alzheimer’s Disease detection using MRI data from the ADNI dataset. The target was to achieve a balanced performance on the test set and to explore the effectiveness of Transformer models in medical imaging analysis.

Model: The implemented Vision Transformer model divides each input MRI image into non-overlapping patches, linearly embeds these patches, and processes them through a series of self-attention layers to learn spatial dependencies. The final output is passed through a classification head that predicts whether a given MRI scan indicates 'Normal' or 'Alzheimer’s Disease'.

Performance: The best model configuration achieved a training accuracy of 67.78%, validation accuracy of 69.66% and test accuracy of 68.20%. The training process included early stopping to avoid overfitting and a detailed analysis of the training and validation metrics.

Dataset Preparation (dataset.py):

• This module defines a custom PyTorch DataLoader that handles the ADNI dataset, resizing images to 224x224 pixels, applying data augmentation, and normalizing pixel values. The dataset is organized into 'train' and 'test' folders, and additional splitting ensures that the training data is further divided into training and validation sets.
• The transformations include random horizontal and vertical flips, sharpness adjustment, and cropping to improve model generalization.

Network Modules (modules.py):

• The Vision Transformer is implemented with a patch embedding layer, positional encoding, multiple Transformer blocks, and a final classification head. The architecture has been modified to suit grayscale MRI images, with appropriate configurations for embedding dimensions and self-attention layers.

Training Script (train.py):

• This script outlines the full training pipeline for the Vision Transformer model. It includes the definition of the optimizer (Adam), learning rate scheduler (StepLR), and training and validation loops. The script also incorporates early stopping to enhance model performance and saves the best model based on validation accuracy.

Testing Script (predict.py):

• This script loads the trained Vision Transformer model and evaluates it on the test dataset. It calculates accuracy, generates a confusion matrix, and saves performance plots. The script highlights key metrics and provides visualizations to aid in assessing model performance.

Performance and Visualizations:

• The model achieved a training accuracy of 67.78%, validation accuracy of 69.66% and test accuracy of 68.20%, with training and validation plots generated to illustrate the learning progress.
• The confusion matrix reveals areas where the model struggled, particularly in identifying false negatives, suggesting potential areas for improvement.

Request:

Kindly review the changes implemented in this pull request. If you find the modifications suitable, consider merging the request. Your feedback on possible improvements or further adjustments would be highly valued.

Thank You
Swastik Lohchab
47415056