Alzheimer's Disease Classification using Graph Attention Networks and LSTMs

Deep Learning GNN GAN LSTMs Medical

Overview

The project aims to classify the severity of Alzheimer’s disease into three distinct stages (Non-Demented, Very Mild Dementia, and Mild/Moderate Dementia). To achieve this, we developed a hybrid deep learning pipeline that treats 2D MRI slices as structured graphs and models the whole MRI scan as a sequential volume:

  • Preprocessing: Individual 2D MRI brain slices are converted into graph structures . The images are segmented into “superpixels,” where each node represents a distinct anatomical region (using features like intensity and variance), and the edges represent their spatial adjacencies.
  • GNN: A Graph Attention Network (GAT) processes each slice’s graph independently. The GAT learns relationships between the brain regions, outputting a rich embedding for each individual slice.
  • Sequential Modelling (LSTM): Since a single brain scan consists of approximately 61 sequential slices, treating them as isolated images loses critical 3D context. An LSTM network takes the sequence of GAT embeddings from a single patient’s scan and processes them sequentially to capture the progressive changes across the brain volume, outputting the final severity classification.

References

Refer to the GitHub repo for more details on the implementation.