Diffeomorphic Iterative Centroid Methods for Template Estimation on Large Datasets

Claire Cury (1) , Joan Alexis Glaunès (2) , Olivier Colliot (1)

1 ARAMIS - Algorithms, models and methods for images and signals of the human brain = Algorithmes, modèles et méthodes pour les images et les signaux du cerveau humain [ICM Paris]
2 MAP5 - UMR 8145 - Mathématiques Appliquées Paris 5

A common approach for analysis of anatomical variability relies on the stimation of a template representative of the population. The Large Deformation Diffeomorphic Metric Mapping is an attractive framework for that purpose. However, template estimation using LDDMM is computationally expensive, which is a limitation for the study of large datasets. This paper presents an iterative method which quickly provides a centroid of the population in the shape space. This centroid can be used as a rough template estimate or as initialization of a template estimation method. The approach is evaluated on datasets of real and synthetic hippocampi segmented from brain MRI. The results show that the centroid is correctly centered within the population and is stable for different orderings of subjects. When used as an initialization, the approach allows to substantially reduce the computation time of template estimation.

Domaines: Imagerie; Géométrie métrique [math.MG]; Traitement des images [eess.IV]; Géométrie différentielle [math.DG]; Imagerie médicale

Claire Cury, Joan Alexis Glaunès, Olivier Colliot. Diffeomorphic Iterative Centroid Methods for Template Estimation on Large Datasets. Frank Nielsen. Geometric Theory of Information, Chapter 10, Springer, pp.273-299, 2014, ⟨10.1007/978-3-319-05317-2_10⟩. ⟨hal-00939326⟩