Computer Vision News - January 2022

64 Medical Imaging Challenge “ One of the things we noticed was people weren’t focusing on tasks that were clinically interesting, ” Keelin points out. “ People were focusing on tasks they have labels for, and some of the labels are things like: Does the patient have a pacemaker inserted? That’s not a clinically interesting question. You can just see it immediately on the image! Lung nodules, by comparison, can be very difficult to see in the chest X-ray because of the projections of the ribs, heart, and diaphragm. If a patient has come in for a heart check, the radiologist may not even be looking for nodules and so could very easily miss one. They’re only human. With AI, we can make suggestions to look more closely at certain parts of the image in case they have missed something. ” to detect nodules. The generation track is for developing algorithms to generate nodules at the requested location. “ We hope the generation track algorithms actually help to improve the performance of the detection track, ” Ecem explains. “ In that way, it’s even more collaborative. We’re all pushing each other. One track is trying to improve the performance of the other track, so we’re all trying to get the best performance together. ” Ecem’s co-supervisor, Bram van Ginneken , is the co-founder of grand-challenge.org . He loves organizing challenges and using the platform to see everything in a fair and comparable way. This challenge came out of his and the team’s excitement around generative modelling.