CNN for MRI Imaging Data

I joined the Neuroinformatics R&D Group at the University of Washington (https://neuroinformatics.uw.edu/)for 3 months to gain experience in a research setting. Over the three months, I worked on developing a custom Convolutional Neural Net model to process Tractometry data. While CNNs are normally used to process image data with the shape of (height, width, channels), with height * width = total # of pixels and channels being the # of color channels, our approach was a little different. Tractometry data is not normal 2D image data, but rather is a mapping of tracts in the brain. We were able to use the notion that channels specify data that has a spatial relationship, and use it to pick up the spatial relationship of bundles of tracts in the brain.

​

Using this logic, I developed a CNN that had state of the art prediction results for the team at 56%! This is very very good for medical imaging data :) After the tuning of the model, I worked on abstracting it into a Scikit-Learn type model class that could be easily used by other neuroscientists.