MICCAI 2021 Daily - Tuesday

“ We create a self-supervised task to give the network a kind of mock practice exam , ” Jeremy tells us. “ We give it examples that we artificially create, introducing subtle irregularities into the input images. Because we’ve introduced them, we know exactly where they are and to what degree they have been altered, and we ask the network to identify that. We try to give it as much practice as we can to make this task as relevant as possible with the aim of generalizing to real pathologies or irregularities . ” In terms of next steps for this work, Jeremy hopes to integrate some adaptive components. The team have submitted an entry to the Medical Out-of-Distribution Analysis (MOOD) Challenge . The idea behind that is to try and use meta-learning to adapt to each new test image. A big part of outlier detection is not knowing what irregularities the test image will contain. “ When you have the new test image, you can’t adapt to it directly because you don’t know if it is normal or abnormal, ” he says. “ You can’t give it a label to adapt to because that label is just based on your assumption . Instead, we’re doing an adaptation using a self-supervised task on the test image. We’re trying to slightly adapt to some of the features present in the test image. The results of the challenge haven’t come out yet, so it’s still a work in progress! ” To learn more about Jeremy’s work [Paper ID 2129], you are invited to visit Poster Session We-S3 Computer Aided Diagnosis tomorrow (Wednesday) at 16:00 – 17:30 UTC. 19 DAILY MICCAI Tuesday Jeremy Tan