MICCAI 2021 Daily – Thursday

To solve this, Spiros and the team proposed to inject noise to encourage local deformations to introduce stochastic variation only at the part that the arithmetic is taking place. “ I know the spatial location, but I want to create some variation so that when I swap out the factor, the new factor will fit in the hole and change the neighboring anatomical factors, ” he explains. “ When I swap in the thicker myocardium, the noise injected at the location of the hole will take care of not allowing the two factors to overlap but create some local deformations so the new factor will fit in a very realistic and plausible way. ” The image shows an example of this model in action. Subject A is healthy, while Subject B has hypertrophic cardiomyopathy . To create a new variant of A that has the same condition as B, the anatomy factors of A and B are extracted and then the myocardium and left ventricle from B are swapped to A. Subject A’ will have the same surrounding anatomy, but a new left ventricle and a new myocardium that are not exactly the same as Subject B. The red arrows show the thick myocardium, and the blue arrows show the local deformation of the surrounding anatomy. The right ventricle has been pushed a bit to the right and the bottom part of the heart has been pushed lower to facilitate this new disease. As well as the disentangled anatomy arithmetic and the noise injection mechanism, which are the two main novelties of this work, another interesting part is the use of two critics to judge how realistic and plausible the generated image is . “ We use a traditional discriminator to be sure that our image has a certain quality and is as realistic as possible to fool the discriminator, ” Spiros explains. “ We also use a pathology classifier that is pre-trained to check 6 DAILY MICCAI Thursday Oral Presentation