2022/03/17 by Fabian Bongratz, Bongratz, Fabian, Anne-Marie Rickmann +5 · 6 citations
Medicine · #Advanced Neuroimaging Techniques and Applications #Advanced MRI Techniques and Applications #Medical Imaging Techniques and Applications
paper · pdf · doi:10.48550/arxiv.2203.09446
The reconstruction of cortical surfaces from brain magnetic resonance imaging\n(MRI) scans is essential for quantitative analyses of cortical thickness and\nsulcal morphology. Although traditional and deep learning-based algorithmic\npipelines exist for this purpose, they have two major drawbacks: lengthy\nruntimes of multiple hours (traditional) or intricate post-processing, such as\nmesh extraction and topology correction (deep learning-based). In this work, we\naddress both of these issues and propose Vox2Cortex, a deep learning-based\nalgorithm that directly yields topologically correct, three-dimensional meshes\nof the boundaries of the cortex. Vox2Cortex leverages convolutional and graph\nconvolutional neural networks to deform an initial template to the densely\nfolded geometry of the cortex represented by an input MRI scan. We show in\nextensive experiments on three brain MRI datasets that our meshes are as\naccurate as the ones reconstructed by state-of-the-art methods in the field,\nwithout the need for time- and resource-intensive post-processing. To\naccurately reconstruct the tightly folded cortex, we work with meshes\ncontaining about 168,000 vertices at test time, scaling deep explicit\nreconstruction methods to a new level.\n