Identifying cortical sulci from localization, shape and local organization

In this paper we propose an approach to identify sulci from sulcal pieces. Our method is founded on the sulci localization, feature-based shapes and their local organization. The position data enable the devising of an easy handled 3D probabilistic atlas using SPAM models. Shapes and local sulci scheme are recognized thanks to SVR models (a regression version of support vector machine). All these aformentioned aspects are merged into a unified Markovian framework, which favours locally the most reliable information. The first model is used to strongly constrain label coverage over space and the second to reach coherence within sulci neighbourhood. The mixture outperforms both models taking the best of their local performances.