MICCAI 2021 Daily – Thursday

This work proposes a data augmentation model to synthesize cardiac images , mixing and matching anatomy factors of variation that are already captured in existing datasets. The work was inspired by a process called vector arithmetic , which is well known in the vision domain. It differs from a traditional synthesis model in that it does not sample from a noise distribution, instead extracted or encoded factors of variation of an image are combined to generate an intermediate representation that preserves the fidelity of the encoded factors. The traditional goal of augmentation is to increase the number of samples, but this method also increases the diversity in a controllable way , so it is more meaningful. It can generate data for parts of the population for which medical data is hard to acquire. The model uses spatial and semantic factors instead of vectors, and the challenging part is mixing and matching spatial factors to present something plausible. Human anatomy has a certain nested structure. You cannot have factors overlapping each other. Lack of plausibility is often what prevents the computer science community from making a contribution through synthesis to the medical community. If results have to be checked one by one, then it is more of a problem than a solution! Spyridon ‘Spiros’ Thermos is a postdoctoral researcher at the University of Edinburgh, working with Sotirios Tsaftaris, and in collaboration with Canon Medical Research Europe, working with Alison O’Neil . The team, together with the Royal Academy of Engineering, have just held a successful tutorial, DREAM: Disentangled Representations for Efficient Algorithms for Medical Data, and have been selected for not one, but two oral presentations at MICCAI this year! Spiros is first author on a paper proposing a novel approach to synthesizing images. He speaks to us about it ahead of his oral presentation today. Controllable cardiac synthesis via disentangled anatomy arithmetic 4 DAILY MICCAI Thursday Oral Presentation