Computer Vision News - July 2019

To solve this, they are leveraging some recent work in medical imaging: Adrian V. Dalca and Guha Balakrishnan ’s VoxelMorph . The idea is that if you have a method that can compute the transformation from your labelled scan to an unlabelled scan, that allows you to compute a set of transformations. Once you have the set of transformations, you apply them back to the labelled scan and use that to synthesise more labelled examples. Co-author Adrian Dalca comments that " This work shows how leveraging unlabeled data and machine learning can lead to tangible and practical improvements. This is important, since the next step is to use this algorithm in practical clinical analyses ." Co-author Guha Balakrishnan adds that “ The benefit of the method is that it is interpretable and a simple idea that can extend to various other data, both in and out of medical imaging ." Thinking about how to develop this work, Amy points out that they’ve demonstrated it on a fairly limited range of scans so far. They just look at T1 MRIs and show what happens if you assume that you have one labelled scan and a hundred unlabelled scans. What happens if you have multiple labelled scans available, or if you apply this approach to different kinds of MRIs or CTs? They think it would work, but don’t know how well, and are really interested to find out. 27 DAILY CVPR Thursday Amy Zhao What happens if you have multiple labelled scans available, or if you apply this approach to different kinds of MRIs or CTs?