Pictorial multi-atlas segmentation of brain MRI
Chengyi Liu, Juan Eugenio Iglesias, Zhuowen Tu · 2012
The use of a single labeled volume (“atlas”) is limited in registration-based segmentation because it is hard for one atlas to represent the whole data population, especially if input images observe large variation. Moreover, the choice of volume to label biases the algorithm. Multi-atlas segmentation has emerged as an alternative but it has a similar drawback due to combinatory combinations of different anatomical structures; and in addition, the computation time grows linearly with the number of atlases. In this paper, a pictorial-structure-based approach to achieving both the high performance and the efficiency of registration-based segmentation is proposed. Our method performs segmentation via registering each structure of the atlas in an exemplar-based graphical model. We compared the proposed approach with multi-atlas segmentation and show the advantage of our method in both effectiveness and efficiency.