A complicating factor in trying to study and ultimately cure illnesses is that it can be uncommon enough to make it difficult to have real data to work from. One example of this is Alzheimer's and brain tumors, but a group of researchers from NVIDIA, the Mayo Clinic, and the MGH & BWH Center for Clinical Data Science applied a type of AI system to produce the desired information. Not only could this lead to improving our understanding of those illnesses, but it could potentially be applied to others.
The researchers used what is called Generative Adversarial Networks (GANs), which has two AI trying to out-compete each other. While one works to generate a synthetic result, the other works to identify it. As this process continues through multiple iterations, both AI networks improve. The researchers fed the GANs real MRI images of people with Alzheimer's and brain tumors with the challenge being to create synthetic versions that are still accurate enough for study. The training also involved the use of labels that would allow researchers to specify characteristics such as the size and location of a tumor.
Another advantage to this method, besides having more data to work from, is that this could be used to anonymize real MRI images, protecting the privacy of patients when any data needs to leave the hospital. More work still needs to be done though, including blind testing to guarantee the quality of the synthetic images.
Back to original news post