Led by Vivek Prabhakaran, MD, PhD, a group of researchers at UW-Madison are investigating ways to generate synthetic positron emission tomography (PET) images from magnetic resonance imaging (MRI) data. By developing a reliable method to “translate” MRI images into certain types of PET images, the group hopes to cut down on the cost and risk for the patient while still being able to accurately diagnose disease.
PET imaging reveals changes in metabolic processes while structural MRI only provides information about brain anatomy. Usually, these imaging modalities are considered complementary and are used together to diagnose neurological diseases like Alzheimer’s disease (AD) and dementia. However, PET imaging involves injecting a radioactive substance into a patient. The ability to convert structural MRI images to PET images (a process called “inter-modality transformation”) could provide a screening tool for further evaluation.
The research team at UW-Madison – which includes Veena Nair, MD and Vikas Singh, PhD alongside Dr. Prabhakaran – is using machine learning to achieve that goal. Their model, DUAL-GLOW, was able to generate sharp synthetic PET images from structural MRI images. These “translated” images, derived from MRI images, were of comparable quality to standard PET images and could be used to identify subjects with AD.
This project was supported by a $75,000 grant from the Boerger Research Fund for Alzheimer’s Disease and Neurocognitive Disorders, which is an initiative of the American Society of Neuroradiology (ASNR) Foundation.
While their project focused specifically on brain images from AD patients, there are numerous potential applications of image-to-image mapping for other diseases, too.
“We are hoping that it will be applied to help diagnose AD,” says Dr. Prabhakaran. “But in the future, we are hoping it can be applied to help with diagnosis for epilepsy or tumors.”
Read more about the project in their published article, “DUAL-GLOW: Conditional Flow-Based Generative Model for Modality Transfer,” here.