43 Anne-Marie Rickmann Computer Vision News This is the task of extracting surface representations of the cerebral cortex from brain MRI scans. Our work Vox2Cortex, which we presented at CVPR 2022, is a combination of a 3D U-Net which learns a voxelwise segmentation of the brain – and a graph convolutional network that takes a template mesh as input and learns vertex vise displacement vectors. Information is passed from the U-net to the graph network to guide the deformation process. The strength of this approach lies in its use of template deformation, maintaining consistent topology and mesh connectivity, thus simplifying group comparisons. Later we presented an extension of this work, V2CC, at MICCAI 2023. V2CC enhances the process by establishing vertex correspondence with the template. This development eliminates the need for cumbersome post-processing registration steps, allowing for straightforward comparisons between subjects or with reference groups. We achieved this by pre-registering training data to the template and optimizing an L1 loss. This seemingly simple strategy has proven highly effective. It also eases applications like comparing cortical thickness in Alzheimer’s patients with healthy controls, which is now much faster and does not require registration. Big thank you to my supervisor Christian Wachinger, for his support throughout the last years and my colleagues at AI-Med! Abdominal MRI scans after gallbladder resection – nnU-Net predicts a hallucination of the gallbladder (white arrow) and HALOS doesn’t. Typical template deformation approach – which does not guarantee vertex correspondence and relies on a post-processing step of registration. Our V2CC approach directly predicts a mesh with correspondence to the template
RkJQdWJsaXNoZXIy NTc3NzU=