A supervised graph-cut deformable model for brain MRI segmentation

© 2013, Springer Science+Business Media Dordrecht. Accurate automatic segmentation of subcortical brain structures in Magnetic Resonance Images (MRI) is of great interest in the analysis of developmental disorders. Segmentation methods based on a single or multiple atlases have been shown to suitably localize brain structures. However, the atlas prior information may not represent the structure of interest correctly. It may therefore be useful to introduce a more flexible technique for accurate segmentation. A fully-automatic segmentation method for brain MRI is considered, which defines a deformable model combining an atlas-based segmentation strategy with a supervised Graph-cut model. The Graph-cut model is adapted to make it suitable for segmenting small and low-contrast brain structures by defining new data and boundary potentials of the energy function. In particular, information concerning the intensity and geometry is exploited, and supervised energies based on contextual brain structures are added. Furthermore, boundary detection is reinforced using a new multi-scale edgeness measure. The method is applied to the segmentation of the brain caudate nucleus in a set of 39 pediatric attention-deficit/hyperactivity disorder (ADHD) patients and 40 control children, as well as in a public database of 18 subjects. We evaluate the quality of the segmentation using several volumetric and voxel by voxel measures, and present a quantitative volumetric analysis of caudate abnormalities in pediatric ADHD. The obtained results show improved performance in terms of segmentation accuracy compared to state-of-the-art approaches.