MICCAI 2019 Tuesday

MICCAI 2019 DAILY Poster Presentation Christian Baumgartner is a postdoctoral fellow at ETH Zürich. He speaks to us ahead of his poster session today. This work is about learning a method that can segment anatomical structures from multiple experts. In medical image analysis, there is often the problem that expert annotations disagree with each other. This method manages to capture this variability of annotations, and after training, you can then sample plausible segmentations rather than just predicting a single one. This inherent uncertainty needs to be reflected because in downstream tasks such as radiotherapy planning it is very important that you know where annotations are certain and where they are uncertain. Using this method, you can analyse the variability of predicted segmentation maps, and say, for example, there is a big difference in that corner, but in that corner, all experts agree. Chr ist ian tel ls us the idea for this work came about because he had been thinking about a closely related work cal led the Probabi l ist ic U-Net . He had some ideas for improving i t , and in the end, fundamental ly rethought theproblem and came up wi th a new solut ion. He is very pleased to add that this new solut ion outperforms the Probabi l istic U-Net paper. In fact , the method outperforms state-of-the-art on mul t iple metr ics. Not only are the numbers bet ter, but also, i t is stat ist ical ly signi f icant . A surpr ising side effect is that , even though they have the ful l probabi l ist ic model , the segmentat ion accuracy does not suffer. In fact , the resul ts are sl ight ly bet ter than for a standard U-Net . He explains the pract ical appl icat ions of the research: “Uncertainty is a very important and widely discussed topic in the communi ty wi th many appl icat ions. I f our vision as a communi ty is to create ful ly automated systems, i t ’s very important that we have quant i f icat ion of uncertainty at di fferent steps so that error propagat ion can be control led. On the one hand i t has appl icat ions in automat ion of steps, but i t also has appl icat ions in di rect ly informing cl inical pract i t ioners about the certainty or uncertainty of an output from a machine learning algor i thm.” 16 PHiSeg: Capturing Uncertainty in Medical Image Segmentation "This inherent uncertainty needs to be reflected"