MIDL Vision 2020

Oral Presentation 12 Richard Shaw is a PhD student at UCL under the supervision of Jorge Cardoso and Carole Sudre. His work is about creating an automated quality control system for MRI. His paper is co-authored with Jorge, Carole and Sebastien Ourselin. Richard speaks to us ahead of his oral and poster session today. The aim of this work is to create a warning system for clinicians and radiographers that will let them know if something affected image quality during a scan so that they can redo it straight away. For instance, if a patient moved, or if themachinemalfunctioned in some way. It does this by detecting movement artefacts. Movement artefacts are notoriously difficult to detect . You have to examine every slice in the volume to find them. This work proposes to complete this task automatically using deep learning . Specifically, it proposes to use uncertainty predictions , where uncertainty can be used as a measure of image quality. In this case, image quality is defined as our ability to perform a task and how uncertain we are in doing that task. The aim is to predict uncertainty using a standard Bayesian neural network model and, by assuming that the sources of uncertainty are independent, decouple it to find out how much of the uncertainty relates to a particular artefact. If there is noise in the image, how much of the uncertainty is due to noise? If the patient has moved, how much of the uncertainty is due to the patient moving? Richa rds tells us he has drawn upon work by Alex Kendall and Yarin Gal on Bayesian uncertainty: “The Bayesian neural network predicts a variance and we have a novel network architecture to split the variance up into independent quantities. It’s tricky because the variance is learned in an unsupervised way. We don’t have labels of uncertainty. It’s learned by maximising A Heteroscedastic Uncertainty Model for Decoupling Sources of MRI Image Quality Richard Shaw