Computer Vision News - September 2019

Challenge 14 The Irontract Challenge 2019 is a collaborative effort between Harvard Medical School, the University of Rochester, and QMENTA Inc. We speak to one of the organisers, Anastasia Yendiki, who is Assistant Professor in Radiology at the Harvard Medical School. Anastasia,what is themainexpected outcome from this challenge? The outcome will be an objective ranking of the accuracy of algorithms for tractography, a technique used to reconstruct the wiring of the brain from diffusion-weighted MRI scans. The ground truth of this wiring is not known in detail - we just know the main highways of the brain but we do not have the complete map of side streets. Although we know that diffusion MRI tractography is reasonably good at following these highways, we also know that it sometimes takes wrong turns. That is because MRI does not image brain pathways directly but instead measures how water molecules diffuse in and around these pathways. Inferring the shape of the pathways from this diffusion is an error-prone process, therefore having access to ground truth is critical. Here we have ground truth from chemical tracer injections in the same brains that the MRI scans come from. What have you learned to date from this challenge, that you can share with our readers? Our preliminary studies on these data have led to some interesting findings. For example, counter towhat onemight have expected, we have found that there are limited gains in tractography accuracy from certain improvements in data quality that the field has been focused on recently. These include increasing the angular resolution (i.e., increasing the number of directions in which water diffusion is measured) or acquiring data at multiple levels of diffusion contrast (i.e., varying a factor called the b-value). This has been the case for the small set of tractography methods that we tested on these data, and it will be interesting to see if the conclusion holds up when we open it to all the other methods that challenge participants will deploy. Anastasia Yendiki The image shows a small fraction (1%) of the paths produced by a typical whole-brain diffusion tractography analysis. Photo by Feda Eid