Computer Vision News - February 2023

39 Sybil: Predict Future Lung Cancer Risk some clever work on neural architectures , and a few trade-offs, the team has finally achieved its goal, but the work does not stop here. Lecia V. Sequist and Florian J. Fintelmann at MGH are at the forefront of the project’s next stage. Analyses and studies are being run in their clinical setting to understand whether Sybil can maintain its performance enough to be used routinely in the real world. “We don’t know the answer yet, so they’re trying to figure that out and, simultaneously, establish some protocols with other collaborators, ” Peter says. “ We’d like more validation on patients not represented in our datasets , which is necessary for it to become something used daily in the clinic. ” The group plans to make its code and models available to the community so that people can analyze them, research the topic, and hopefully produce even better models in the future. “ As a group, we’re going more into chemistry and biologics, trying to model proteins and small molecules, ” Jeremy reveals. “ We’re moving a little bit away from imaging and diagnostics and are trying to use AI to accelerate therapeutic development . I think that’s the next big frontier for us! ” accessing it, managing the collaborations, and dealing with the review process. That takes a long time. ” Early on, the group realized they needed doctors to tell them if cancer was present in the low-dose CT scans and where it was. It then took some time to complete the manual task of labeling hundreds of scans for training . The teamruledout using clinical risk factors in the model and found that combining multiple CTs for a patient over time, as a doctor would do in a clinical setting, put too much strain on the system. In the end, the group’s most significant challenge came down to one thing: scale . “ When you work with super large images, you have potentially hundreds of slices, and with these huge volumes, things can get heavy computationally , ” Jeremy explains. “ We hit many roadblocks with just being able to run experiments and iterate fast enough. We were lucky to have bigger, better machines with more state- of-the-art GPUs over time. We’d be in a much better situation if we were to start the project today, as our compute at the beginning was not even remotely close to where it is now. ” Nonetheless, with much engineering,