“We’re trying to make the task easier by having a model that you just need to use instead of build,” Jose explains. “Then, clinical researchers will have a much easier time just using the model. Instead of months, hopefully, it will take them hours or minutes. Beyond that, it will help them iterate faster on their research. We’re still doing the same segmentation tasks; we’re just making the whole process simpler to reach a wider audience of researchers.” When we ask whose idea UniverSeg was, the team quickly emphasize that it was a collaborative effort spanning multiple institutions over several years, including Cornell, MIT, and Harvard. It involved extensive work on data, training on over 50 datasets, and iterating on what is a consistent format for the model to train on. Decisions on how to evaluate on out-of-distribution datasets and architectural choices, such as the new CrossBlock mechanism introduced to produce accurate segmentation maps, required the combined effort of researchers to create a model capable of adapting to new tasks seamlessly. While UniverSeg is a significant boon to research, its potential extends to real-world medical applications. Jose acknowledges that deploying such technology in the medical field, with its high stakes, is challenging. “If you had told me this was possible when I started my PhD, I would not have believed you,” he reveals. “This is a really good first milestone. As people keep working on it and improving on the ideas, I see a bigger potential for getting this into real-world systems, but it requires a lot more testing and evaluation whenyou’re making very important decisions.” Victor continues: “A large part of our work is helping boost the efficiency of these medical professionals. Suppose you have a small dataset tied to a particular disease or sub-group of people you don’t have a model for. You’d like to measure things about the subjects’ brains, such as white matter hyperintensities or hemorrhage, specific to this one group. Segmentations of these quantities 9 DAILY ICCV Friday UniverSeg
RkJQdWJsaXNoZXIy NTc3NzU=