I wanted to develop a diagnosis tool for doctors to envision how a patients health might evolve and decided to focus that tool on brain health. Getting a brain scan is an expensive time consuming process so I wanted to make it easier to visualize what might be occuring.
What it does
Encodes the image and age representation of an MRI in a Variational Autoencoder and then at inference time samples different ages to see how the brain would look in the future.
How we built it
Using the Pyro VI library, pytorch and a public brain database we built a model to understand how the brain ages. The model trains itself to autoencode the brain and predict the age and then later on we can impute the age to see how the brain would change when reconstructed.
Challenges we ran into
The model does not output images at a high enough resolution which is a common problem with VAEs. Training also took much longer than expected and I canceled a few trial runs because I thought it wasn't working but it required training overnight.
Accomplishments that we're proud of
We developed a working model which shows some signs of brain aging.
What we learned
The pyro AI library for VI.
What's next for Brain Aging Visualizer
Building in more latent variables and incorporating different views of the brain to refine its image outputs would be a nice next step. Hopefully we could even add in treatments so that patients can see the positive outcomes on their own brain. More importantly, a UI so it can be deployed within an MR machine.