MIDL Vision 2020

13 Richard Shaw the log likelihood. We use a sequence of networks to predict pseudo labels of uncertainty. We first have a network that predicts the segmentation and the uncertainty of only clean images. They’re our pseudo labels. We then freeze that network and train a new network for a particular artefact. For example, the network sees noisy images and it predicts an uncertainty. We use the uncertainty from the first network as a supervising label to be able to split the uncertainty apart. It’s like a series of student-teacher networks where the network before teaches the network after to decouple uncertainty.” Classical methods of measuring image quality are quite simplistic, tending to just estimate the signal-to-noise ratio between slices, for example. Learning from the data is better , and since theuncertainty is learned from the data in an unsupervised way, that is particularly beneficial. “I really enjoy working with all of them!” Augmentations Nnetwork-architecture