Author: Noah James     Major: Computer Science

Photo of Noah James

My name is Noah James, and I am a Computer Science student. Starting in the summer of 2020, I led a unique bioinformatics and machine learning research project under the guidance of Dr. Justin Zhan’s Data Science and Artificial Intelligence lab as well as mentors at the Air Force Research Laboratory at Wright Patterson Air Force Base in Dayton, Ohio. We modified our goal as we progressed through the research, but our primary objective from the beginning was to increase the accuracy of brain disease diagnosis from magnetic resonance images (MRIs) with a single machine learning algorithm. For testing, we used an open-source dataset of individuals with and without spinocerebellar ataxia type 2 (SCA2), a debilitating neurodegenerative disease. If our algorithm could detect the disease even in its early stages, then in practical application patients could begin receiving treatment when it is most important. To accomplish this, we explored the merging of MRIs with a similar data type known as diffusion tensor images (DTIs) in order to enhance diagnostic accuracy.

My research in the past has been about neural networks and considering that they receive the majority of attention, as well as repeated groundbreaking results as new variations, are invented, we decided to base our machine learning model on them. Recent literature on fusing MRIs and DTIs described how new DTIs could be created with the use of generative adversarial networks (GANs), a unique type of generative neural network architecture developed in 2014 that involves two opposing neural networks. Even further, a new extension from 2017 known as CycleGAN allowed for the learning of a translation between two imaging modalities, like in this case. We decided to start from there, coding a new CycleGAN to first translate input MRIs into DTIs. Why would we need DTIs though?

A DTI shows the rate of diffusion of water molecules everywhere in the brain. We use a 3×3 matrix known as the diffusion tensor to represent the diffusion in the 3 cardinal directions at every voxel of a DTI. Most importantly, this diffusion will show where there is damage in the white matter of the brain. Since neurodegenerative diseases like SCA2 generally cause this sort of damage, DTIs will help us pinpoint it so we can accurately diagnose the disease. However, creating high-resolution DTIs is actually much more difficult than getting an MRI. Most people never get DTIs of their brains, so we used the aforementioned CycleGAN to generate a high-resolution, realistic DTI from a patient’s MRI. This synthetic DTI becomes part of the basis for our computer’s diagnosis.

The power of neural networks comes into play here. To describe it rather simply, a neural network is just a bunch of multiplications and additions of whatever data is received as an input with each operation being multiplied by a learned ‘weight’. The output of a neural network can thus be represented by a single mathematical function. For really large networks, this function can be really complicated, but fortunately, we have the power of computers, which can make billions of calculations a second, at our command. As long as we can derive an appropriate mathematical metric for the overall network performance, known as the ‘loss’ function, we can use a technique known as backpropagation to adjust the weights of the network neurons, thereby building the best function to accomplish whatever task we have set it up for. In this case, the function turns a collection of MRI voxels into DTI voxels. After learning this function with the CycleGAN network, we put together another neural network for the actual diagnosis task. This network consisted of two parallel residual networks, each making its own prediction of either patient or control. These predictions were averaged for the overall diagnosis. After testing our method, we found that it could garner 100% accuracy on the SCA2 dataset. Our method turned out to be a success.

We initially decided to pursue this line of research because of the group’s interest in medical imaging and the biomedical field in general. I had been working in Dr. Zhan’s lab previously, where I had worked on computer vision research, so the transfer to medical imaging was not too jarring. However, from the beginning, the entire research process was extremely difficult. The mathematics of the methodology alone is exceedingly complicated, not to mention the actual coding of it. At one point, we came across the issue of MRIs draining memory resources very easily. Just one MRI, comprising about 1 gigabyte of computer memory, contained about 32 million voxels to be saved to the graphics processing unit cache, and the nature of neural networks causes intermediate calculations to be saved, effectively blowing this number up even more. We regularly had excessive memory allocations that crashed the program, so I had to come up with a series of algorithms in order to reduce the memory usage. After having succeeded, I am grateful for the learning experience of working with high-volume data. I am also grateful for the mentorship I received from Dr. Zhan and the mentors at the Air Force Research Laboratory. Finally, I am grateful for the expertise and help I received from fellow students Theresa Edattukaren and James Roddy in pursuing this research. In the future, I expect to continue to research novel machine learning algorithms to advance computer vision.