The inspiration came more from wanting to use TensorFlow probabilities. We believe that giving doctors and researchers access to the specific probabilities of a model's outputs will allow for them to make better decisions. Understanding this, we wanted to build a proof of concept that shows a model going a step further: not just outputting a value representing the least loss for an input, but also something that tells a researcher how accurate their final prediction was.
Disclaimer: This is not a production ready model. It is wrong to claim that this is an accurate solution for predicting Parkinson's from an MRI scan. From the data and results, this model has not found anything significant. This project is more so a proof of concept and an application example.
What it does
It takes an MRI image and then outputs the chance that the brain is suffering from Parkinson's and the chance that the neural network is correct.
How I built it
It was written in python. Uses tensorflow 2.0 nightly for eager executions and tensorflow probabilities.
Challenges I ran into
As anyone knows with data, just formatting it to input into the model is always a challenge. Also because the data is bigger, tensorflow doesn't allow for a tensor larger than 2GB in tfp so I had to vastly scale down the batch size.
Accomplishments that I'm proud of
The heatmaps were a nice addition to help visualize what the network sees.
What I learned
Got an understanding of the use cases for tensorflow probabilities.
What's next for Parkinsons 3DCNN Bayesian Network
Fixing it, adding a front end client to have users interact with it.