ATLANTA—Georgia State University researchers, with colleagues at the Massachusetts Institute of Technology (MIT) and the Massachusetts General Hospital (MGH), have received a $2.5 million grant from the National Institutes of Health’s Brain Research through Advancing Innovative Neurotechnologies (BRAIN) Initiative for research aimed at revolutionizing scientists’ understanding of the human brain.
The team will develop Nobrainer, an open-source deep learning framework for 3D image processing, to integrate machine learning into neuroimaging research and clinical applications.
“Advances in artificial intelligence and deep learning can help researchers extract greater insight from brain scans while cutting down on the time it takes to process this data,” said Sergey Plis, associate professor of computer science at Georgia State and institutional lead on the grant proposal. “For example, we could learn more about the specifics of how mental disorders or aging affect the structure of the brain.”
Models that can elucidate these kinds of complex patterns are data-hungry, and assembling huge sets of brain data is challenging, particularly for small research groups.
“When Google wants to create a chatbot, they can train it using data from every Internet search,” said Plis, who is also director of machine learning core at the Center for Translational Research in Neuroimaging and Data Science. “For brain imagers, though, the barriers can be too high. Gathering thousands of brain scans as well as the hardware needed for training is expensive and you have to anonymize the data in order to get around the privacy issues.”
The team is led by Satrajit Ghosh of MIT, Bruce Fischl of MGH and Plis. They plan to create deep neural networks that have been pre-trained on brain scans from more than 65,000 individuals. They will disseminate the technology as a set of widely available tools and ready-to-use models for neuroscientists. The tools and the resulting models will be standardized, ensuring scientists can get comparable results and share them more easily, without patient confidentiality concerns.
The team is developing a unique feature in which the models can critique what they know, quantifying the degree of uncertainty in their own analyses and reporting where they’re likely to be wrong. This could help scientists decide when to trust the model and when more data need to be collected. As more researchers use the models, posing new questions or tuning the models to new datasets, the tools will continue to learn, becoming more accurate.
“The further the model is passed on, like a hot potato, the better it becomes,” Plis said. “When you’re studying something that affects the brain in unpredictable ways, like a stroke, you need a lot of data because there’s a lot of variability in patient outcomes. With the hot-potato style of learning, the model gradually absorbs this variability and becomes better at making predictions.”
Another major benefit is the tools’ ability to process data much faster than available models. The research team trained Nobrainer to make some of the same predictions as Freesurfer, a best-in-class MRI analysis tool developed by MGH. Preliminary studies show the technology outperforms Freesurfer, making some of the same calculations in minutes versus hours. The team plans to work on automating and speeding up other parts of the Freesurfer platform and other types of neuroimaging analysis using their tools. Reducing the time needed to perform complex analytics may quicken scientific and clinical discoveries about the brain.
With a background in engineering, artificial intelligence and computer science, Plis is focused on developing computational instruments that enable knowledge extraction from observational multimodal data collected at different temporal and spatial scales.