Computer Vision News - January 2023

36 SMILE Lab brain/bio-inspired AI Skylar points out that the segmentation output of many of the tools available is part of an overall pipeline that can take 11 hours or more to run. These tools are not particularly user-friendly, and someone unfamiliar with themmay be unable to separate the specific segmentation approach from the rest and would have to wait the total time for each subject to get the output they need. Three seconds is clearly a massive improvement on that. “ The segmentation pipeline is used to get the electrical flow model, which doesn’t sound like something that’s directly translatable to practice, but it’s a crucial part of non-invasive brain stimulation , ” Skylar explains. “ The goal is to have more accurate parameters for non-invasive brain stimulation, so we can use this to treat people with conditions like Alzheimer’s. Not all heads are the same, so you want to give someone the best parameters possible to get the treatment that will most help them. ” One of the biggest challenges in head segmentation research matches one of the biggest challenges in deep learning: a model often struggles with data that is different from the data on which it was trained . This study uses data from two locations, and where it would train on data from one location, it was robust, reliable, and trustworthy AI model that can segment the T1 MRI volume in only three seconds . “ A challenge we face is that our patient population is over 50 years old, whereas public segmentation tools have been built on younger adults around 20 to 30 years old, ” Skylar tells us. “ These public tools don’t account for the natural processes that occur with aging, such as white matter and gray matter degeneration in older adults. We want a tool that can work better for older adults and better for individualized cases since the human head is highly variable. ”