CNN-based segmentation with a semi-supervised approach for automatic cortical sulci recognition

Léonie Borne, Jean-François Mangin, Denis Rivière

Keywords: CNN, segmentation, semi-supervision, cortical sulci

Abstract: Despite the impressive results of deep learning models in computer vision, these techniques have difficulty achieving such high performance in medical imaging. Indeed, two challenges are inherent in this domain: the rarity of labelled images, while deep learning methods are known to be extremely data intensive, and the large size of images, generally in 3D, which considerably increases the need for computing power. To overcome these two challenges, we choose to use a simple CNN that tries to classify the central voxel of a 3D patch given to it as an input, while exploiting a large unlabelled database for pretraining. Thus, the use of patches limits the size of the neural network and the introduction of unlabelled images increases the amount of data used to feed the network. This semi-supervised approach is applied to the recognition of the cortical sulci: this problem is particularly challenging because it contains as many structures to be recognized as labelled subjects, i.e. only about sixty, and these structures are extremely variable. The results show a significant improvement compared to the BrainVISA model, the most used sulcus recognition toolbox.

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