MIDL Vision 2022

8 VISION MIDL The main contribution of this model is that it can flag data as out of distribution in both the near and very far out-of-distribution domains. It can flag slightly corrupted CTs as out of distribution, but also images that are not of the head and not of CTs as out of distribution. “ In my experience of reading the literature and experimenting with other models, they rarely succeed in both domains, ” Mark points out. “ We were very interested in ensuring we covered both, so we built a benchmark dataset of images in both domains. Something I haven’t seen in the literature is that people test their uncertainty or out-of-distribution methods in both domains. I think it’s important to test and that we put extra effort into assessing. ” Thinking about the next steps for this work, Mark points to the map you get from a transformer, which can be reshaped to give a map of where the anomaly is. However, the map is very low-resolution because images are compressed so much. The team is keen to get a high-resolution map of where the anomalies are, either by using less compression or different classes of models, such as diffusion modeling, to do similar things . Also, they would like to combine the uncertainties from this out-of-distribution measure with uncertainties from a task-specific network, such as a downstream segmentation network, to get richer measures of uncertainty from the networks. This work sports a long and impressive list of authors thanks to a grant which is a collaboration between KCL and UCL aiming to develop models that can be deployed clinically. It has been spearheaded by Sebastien Ourselin and M. Jorge Cardoso from KCL and Parashkev Nachev from UCL. “ It can be difficult to work so closely between two teams at different universities, ” Mark reveals. “ Regular meetings and staying in touch asynchronously with Teams or Slack is crucial. Also, having a large working group and then smaller groups collaborating more closely across sites. It’s difficult but rewarding because it forces you to test and develop algorithms on data from different sites and to check that things you develop are robust. ” Mark assures us that he is very keen to open source the code for this work , having received several requests for it. Another paper is currently using the same code base, but once that is out, it will be available to download. Come along to Poster Session 2.1 today, onsite at 15:20 - 16:20 and virtual at 11:00 - 12:00, and Oral Session 3.1: Trustworthy AI tomorrow at 09:40 - 10:40 (UTC+2). Oral Presentation