Sub-cortical brain structure segmentation in Magnetic Resonance Images (MRI)has attracted the interest of the research community for a long time becausemorphological changes in these structures are related to differentneurodegenerative disorders. However, manual segmentation of these structurescan be tedious and prone to variability, highlighting the need for robustautomated segmentation methods. In this paper, we present a novel convolutionalneural network based approach for accurate segmentation of the sub-corticalbrain structures that combines both convolutional and prior spatial featuresfor improving the segmentation accuracy. In order to increase the accuracy ofthe automated segmentation, we propose to train the network using a restrictedsample selection to force the network to learn the most difficult parts of thestructures. We evaluate the accuracy of the proposed method on the publicMICCAI 2012 challenge and IBSR 18 datasets, comparing it with differentavailable state-of-the-art methods and other recently proposed deep learningapproaches. On the MICCAI 2012 dataset, our method shows an excellentperformance comparable to the best challenge participant strategy, whileperforming significantly better than state-of-the-art techniques such asFreeSurfer and FIRST. On the IBSR 18 dataset, our method also exhibits asignificant increase in the performance with respect to not only FreeSurfer andFIRST, but also comparable or better results than other recent deep learningapproaches. Moreover, our experiments show that both the addition of thespatial priors and the restricted sampling strategy have a significant effecton the accuracy of the proposed method. In order to encourage thereproducibility and the use of the proposed method, a public version of ourapproach is available to download for the neuroimaging community.
展开▼
