Computer Vision News - May 2021

20 Sandy Engelhardt is an Assistant Professor at University Hospital Heidelberg, where she leads the Artificial Intelligence in Cardiovascular Medicine research group. The group are organizing the AdaptOR challenge, a deep generative model challenge for domain adaptation in surgery – hence ‘OR’ for Operating Room. It will publish its results during the first Deep Generative Model (DGM4MICCAI) workshop at MICCAI 2021 in September. Sandy tells us more. A key topic of the AdaptOR challenge is to use a deep generative model to transfer images from a surgical simulator and transform them into a more realistic representation. Data provided will include a small number of datasets from real surgeries, and a larger annotated dataset from the simulator, which was developed by Sandy’s department. “To assess the performance of this simulated data, we came up with the idea to pose an additional task on these datasets to automatically detect a specific point in the images in the intraoperative domain , ” Sandy tells us. “When you do this domain translation, you can use these surgical datasets and the intraoperative datasets to create a better performing point detector for the intraoperative domain. It’s a mixture of different kinds of tasks, I would say.” The image-to-image translation topic was first propelled into the mainstream in 2018 following Zhu et al.'s research, which is also the year the team had their first paper on the subject accepted at MICCAI. Since then, they have used their datasets for various tasks, and other researchers have been clamoring for it to be available open source. This will be the first chance they have to work with it. After the challenge, the data will be made publicly available for non- commercial use. Sandy hopes lots of people join in and Sandy Engelhardt MICCAI Challenge