MICCAI 2021 Daily – Thursday

that the generated image has the pathology that we want it to have. If we add a myocardium and a left ventricle that corresponds to a certain condition A, we don’t want the generated image to be classified as condition B because this is a random generation. If we can’t guarantee that the target pathology is there, then we can’t speak about plausibility. ” One final breakthrough is regarding editing. Once the model is trained, the images can be edited by selecting a factor of variation, the left ventricle for example, to change its size and create variations of the same patient with small differences step by step. This offers editing of the anatomy through changing the factor values and opens up new opportunities for image editing in the medical domain . You can see on the left side of the image that the left ventricle is being eroded and creates different versions of the same patient. You can be sure that the generated result is plausible because the data is biased. Every image that the model has seen is a real image and it has learned the smallest and biggest possible left ventricle that can exist. The model will stop increasing the volume of the left ventricle when it hits a specific size. This work is part of a wider project called Healthcare AI , which is aiming to translate disentanglement, domain generalization, causality, and other research parts that the team are exploring into an automated estimation of cardiac biomarkers, disease diagnosis, and cardiac episode prediction. To learn more about Spyridon’s work [Paper ID 527], you are invited to visit his oral during Session Th-Oral-AM-B Image Synthesis today at 11:00 – 12:30 UTC. 7 DAILY MICCAI Thursday Spiros Thermos