CNN-based MRI Image Classification w/ GradCAM Visualization

Inspiration

Alzheimer's diseases is a progressive neurological disorder where early diagnosis is crucial but often challenging. We want to make this project as our first hands-on experience in understanding medical-related needs and knowledge that can be applied with AI, as its uses have seen great benefits in medical field.

What it does

This project classifies brain MRI images into four Alzheimer’s disease stages, No Impairment, Very Mild, Mild, and Moderate Impairment, using a convolutional neural network with transfer learning. In addition to prediction, the model provides visual explanations using Grad-CAM, highlighting regions of the MRI that most influenced each classification.

How we built it

Used a ConvNeXt-Base model pretrained on ImageNet as the backbone
Applied transfer learning, training only the final classification layer
Preprocessed MRI images with resizing, normalization, and data augmentation
Split the dataset into training, validation, and test sets using stratified sampling
Evaluated performance using accuracy and loss metrics
Applied Grad-CAM to visualize spatial attention and improve interpretability

Challenges we ran into

Computational constraints limited training to a small number of epochs due to runtime restrictions
A noticeable generalization gap appeared between validation and test accuracy
Interpreting medical images requires caution, as visual explanations do not guarantee clinical correctness

Accomplishments that we're proud of

Successfully integrating explainable AI (Grad-CAM) into a medical image classification pipeline
Maintaining proper dataset separation to avoid data leakage
Building a reproducible and interpretable deep learning workflow under limited compute resources

What we learned

Transfer learning is effective for medical imaging but requires careful evaluation
Able to learn and utilize new interpretability tools like Grad-CAM are essential for understanding and trusting model behavior
Computational limitations can significantly affect convergence and generalization