Combining 3D U-Net and bottom-up geometric constraints for automatic cortical sulci recognition

Léonie Borne, Denis Rivière, Jean-François Mangin

Keywords: U-Net, geometric constraints, cortical sulci

Abstract: While the limits of deep learning are still to be clarified, some problems may benefit from a mixed approach combining CNNs with traditional strategies. For instance, bottom-up representations embedding domain-specific knowledge could help to regularise a voxelwise segmentation. In this paper, we propose such an approach dedicated to the automatic recognition of the human cortical sulci, designed as the labelling of the voxels of a skeleton of the fluid surrounding the brain. Deep learning is used to provide a top-down perspective to a classical bottom-up pattern recognition system. Our original approach is compared with the approach proposed in the BrainVISA package (www.brainvisa.info), the most used sulcus recognition toolbox. As far as we know, this is the first time that CNNs is used for sulcus recognition. We show that our approach outperforms the BrainVISA method.

Code Of Conduct: I have read and accept the code of conduct.

Remove If Rejected: Remove submission from public view if paper is rejected.