MIDL Vision 2022

Transformer-based out-of-distribution detection for clinically safe segmentation Mark Graham is a postdoc at King’s College London. His paper on working with neural networks out of distribution will be presented as a poster session today and an oral tomorrow; he speaks to us ahead of its debut . Neural networks work extremely well at narrow tasks, such as when trained to do a task on a dataset and then performing inference on data within the same distribution. What they are not so good at is working out of distribution . If a network is trained to segment tumors with MRI, for example, and then given a different modality, it will not do a good job. This paper looks at how to get around this problem in a hospital scenario with many different data types and a network deployed to perform a specific task. In the tumor segmentation scenario , how will the network react if it is presented with data from a distribution it was not trained on? “ This is relevant in a hospital setting because that happens all the time, ” Mark tells us. “ Hospital data is messy and mislabeled . Something can 4 VISION MIDL Oral Presentation