Cortical sulci recognition and spatial normalization

Brain mapping techniques pair similar anatomical information across individuals. In this context, spatial normalization is mainly used to reduce inter-subject differences to improve comparisons. These techniques may benefit from anatomically identified landmarks useful to drive the registration. Automatic labeling, classification or segmentation techniques provide such labels. Most of these approaches depend strongly on normalization, as much as normalization depends on landmark accuracy. We propose in this paper a coherent Bayesian framework to automatically identify approximately 60 sulcal labels per hemisphere based on a probabilistic atlas (a mixture of spam models: Statistical Probabilistic Anatomy Map) estimating simultaneously normalization parameters. This way, the labelization method provides also with no extra computational costs a new automatically constrained registration of sulcal structures. We have limited our study to global affine and piecewise affine registration. The suggested global affine approach outperforms significantly standard affine intensity-based normalization techniques in term of sulci alignments. Further, by combining global and local joint labeling, a final mean recognition rate of 86% has been obtained with much more reliable labeling posterior probabilities. The different methods described in this paper have been integrated since the release version 3.2.1 of the BrainVISA software platform (Riviére et al., 2009).

Accurate identification of 125 cortical structures is achieved through the modeling of inter-subject anatomical variability by coupling localization information with spatial normalization. Figure: resulting probabilistic atlas.Figure optionsDownload high-quality image (95 K)Download as PowerPoint slideHighlights
► We labeled 125 cortical structures on 62 healthy subjects.
► We introduce a joint sulci labeling and registration Bayesian framework.
► Results have been significantly improved with a recognition rate of 86%.
► All proposed models have been released in Brainvisa 3.2.1.