Generating Varied MRI Data with GANs

Coming from a neuroscience background, we were interested in the intersection of brain imaging and machine learning. We’re also aware of the reproducibility crisis, and in particular, the fact that many machine learning models do not generalize well when applied to other data sets. For example, in MRI, the place where the scan is taken can influence many details of the image. We were also very curious about GANs and how they can generate images. So why not use GANs to translate a scan from one site to another? Learn how to translate a brain scan from hospital A to the same brain scan if it were taken in hospital B?

What it does

So far, we were able to load the data and prepare the model structures.

How we built it

Using tensorflow, heavily influenced by the following tutorial: https://www.tensorflow.org/tutorials/generative/pix2pix

Challenges we ran into

The main challenge we faced was our lack of knowledge in deep learning. For that reason, the first hurdle we faced was converting the neuroimaging data into a usable format. For that, we used tensors to try to capture the different images onto a grey-scale format.

Another key technical problem was trying to import Python libraries from the Compute Canada cluster. Commonly used packages such as os were unavailable which required us to workaround outputting files using the pathlib library. This also prevented us from using matplotlib due to incompatibility in the path dependency.

Accomplishments that we're proud of

Understanding what a tensor is :)

Using tensorflow for the first time!

What we learned

We learned some general basics of deep learning, in particular, what a GAN model is and how and why it is used in the realm of machine learning. Furthermore, we were exposed to many different libraries such as tensorflow that allowed us to gain some valuable knowledge in tensors and the role they play in neuroimaging.