MICCAI 2022 Daily – Tuesday

“ In that 3D space, the slices are well aligned with each other after motion correction . We then put the stack of images into a super- resolution algorithm for 3D reconstruction and get the final 3D volume. ” Previous solutions involved traditional optimization-based registration methods to maximize the similarity between the 2D slices. These methods are often ill-posed and suffer from local minima problems. It is difficult to find a global minimum using these methods. “ Later, people used deep learning-based methods, but the approach was kind of naïve, ” Junshen tells us. “ For each 2D image, they put it into a convolutional neural network and independently used the network to predict the position of that slice in the 3D space. In real acquisition, slices are acquired consecutively. ” If two slices are acquired adjacently, they should have a similar motion. Even though the fetal movement is random, it is still continuous, meaning adjacent slices should have a high correlation in their motion . Knowing this gave Junshen the idea to explore the correlations between different slices, and he decided to model the 2D slices as a sequence of images. 5 DAILY MICCAI Tuesday Junshen Xu