A hybrid approach for optimal automatic segmentation of White Matter tracts in HARDI

Whole brain tractography generates a very huge dataset composed by various tracts of different shapes, lengths, positions. Then clustering them into anatomically meaningful bundles is a challenge. Until now, several clustering methods have been proposed such as methods based on similarity measures or methods based on anatomical information, but no optimal clustering criteria were found yet. All methods have deficiencies. The combination of appropriate and distinguishable aspects of both methods is recommended to improve the results. Therefore, the aim of this study was to develop a new combined approach that incorporates various features in order on one hand to overcome the size and the complexity of the tractography datasets and the other for more efficacy and precision of clustering results. We propose a hybrid approach for automatic segmentation of White Matter (WM) tracts in HARDI that combines two complementary levels. The first level of our contribution is based on a similarity measure which aims to reduce the dimensionality of the data. The second level embeds a priori knowledge represented by a subject bundle atlas constructed in this work to improve the result of the clustering. Our method is able to well extract 13 major WM bundles. The results accuracy are measured by a Kappa analysis between the proposed method results and bundle atlas. The average Kappa values is superior to 0.70, it suggests substantial agreement.