Computer Vision News - August 2022

35 Learning Shape Reconstruction... annotations required to establish dense correspondence between the training shapes. Nowadays, there are plenty of segmentations because the segmentation process can be automated, and these shape prior models can be built automatically – even from sparse data, which is the contribution of this work. “ These days, the segmentation task often performs well enough, ” Tamaz points out. “ We have the nnU-Net . We have the contrastive self-supervised pre-training on large unlabeled data sets. We’re going backward and saying, okay, now we have lots of segmentations and shapes, can we learn a probabilistic generative model of those shapes or a shape prior? ” Reflecting on his time at MIDL, Tamaz leaves us with an anecdote. “ It was funny that the conference is called Medical Imaging with Deep Learning, ” he laughs. “ Inmywork, Ihaven’tusedasinglemedical image! I only work with segmentations because it’s about shape reconstruction and building shape priors, which is a very important and useful topic in medical image processing. In the end, what do we do the segmentation for? It’s to obtain the shape of an anatomical structure. I hope that our work attractsmore people into the shape analysis world! ”