Computer Vision News - May 2018

The focus of Ke’s work is body part regression. Ke explains: “ We know that each CT volume contains multiple 2D CT slices. I want to answer the question – given a 2D slice, where does the image come from? Which body part? ” To answer this question, existing methods split the body into several predefined parts – like lung, liver, kidney, etc. – but the application of this method is limited. Ke aims to train a regressor instead of a classifier to predict a value for each slice. Ke feels that this method provides a better answer because it allows a continuous value for each part of the body. To retrieve any part of the body, you can use this value. Other methods are very different and can only achieve this in a predefined way; for example, they can only retrieve the liver or the kidney, but not another part in the middle. What’s particularly interesting about this work is that it only requires unlabeled volumes to train, so it is very easy to collect the training sample. Ke and his colleagues have put their code and training model on Github, so everyone can download it and use it in their projects. Ke tells us that the most challenging part of the work was coming up with the idea itself: “ We designed a self- supervised method, which is like a self- organizing process. We used a slice order information to train our algorithms. This method is very creative, and after thinking of this method we can just write programs to do it and it’s easy, so I think the most important part of our method is to get the idea. ” 22 Friday ISBI DAILY Oral Presentation: Unsupervised Body Part Regression Via Spatially Self-Ordering CNNs Ke Yan Ke Yan is a postdoc from the National Institute of Health, supervised by Ronald Summers . “We designed a self- supervised method, which is like a self-organizing process. We used a slice order information to train our algorithms.” BEST OF