MRI_SOMVAE

This repo contains the code for a Self-Organized Map (SOM) VAE which was applied to T1 weighted MRI brain images from healthy controls in the OASIS dataset. The goal of the project is to find embeddings for local brain regions. The SOM is a 2D grid consisting of nodes organized in a 16 x 16 grid. Each node has a 32 channel vector associated with it. After passing through the encoder convolutional layers, 32 channel voxels in the image are mapped to the best matching unit in the grid based on distance. This node and the nodes 1 away in the grid are updated via backpropagation to become closer to the incoming voxel. Over time, the topological features of the image are captured by the SOM grid. We successfully apply the SOMVAE and find that voxels closer together in the input image are mapped to the same or nearby nodes on the grid.