Segmentation of image ensembles
Spatial priors, such as probabilistic atlases, play an important role in MRI segmentation. However, the availability of comprehensive, reliable and suitable manual segmentations for atlas construction is limited. We therefore propose a method for joint segmentation of corresponding regions of interest in a collection of aligned images that does not require labeled training data. Instead, a latent atlas, initialized by at most a single manual segmentation, is inferred from the evolving segmentations of the ensemble. The algorithm is based on probabilistic principles but is solved using partial differential equations (PDEs) and energy minimization criteria. We evaluate the method on two datasets, segmenting subcortical and cortical structures in a multi-subject study and extracting brain tumors in a single-subject multi-modal longitudinal experiment. We compare the segmentation results to manual segmentations, when those exist, and to the results of a state-of-the-art atlas-based segmentation method. The quality of the results supports the latent atlas as a promising alternative when existing atlases are not compatible with the images to be segmented. Volumes acquired via medical imaging modalities, such as MRI, are frequently subject to low signal-to-noise ratio, bias ﬁeld and partial volume effects. These artifacts, together with the naturally low contrast between image intensities of some neighboring structures, make the automatic analysis of clinical images a challenging problem. Probabilistic atlases, typically generated from comprehensive sets of manually labeled examples, facilitate the analysis by providing statistical priors for tissue classiﬁcation and structure segmentation (Ashburner and Friston, 2005; Fischl et al., 2002; Pohl et al., 2006; Pohl et al., 2007a; Van Leemput et al., 1999). Yet, the limited availability of training examples that are compatible with the images to be segmented renders the atlas-based approaches impractical in many cases. While brain atlases of healthy human adult anatomy are widespread, reliable manual segmentations of newborn brains or of different body regions are not as common. Moreover, in the presence of pathologies where the diversity in structure and appearance is unpredictable, incorporating priors obtained from different subjects is error prone. Recently, a few methods have been proposed to reduce or avoid the dependency on possibly incompatible atlases. Bazin and Pham (2007) proposed an atlas-based segmentation method that uses topological constraints to avoid possible bias introduced by the atlas. Yang and Duncan (2004) employed manually labeled structures to support automatic segmentation of neighboring structures within the same image. Tu et al. (2008) proposed a discriminative approach for segmentation of adjacent brain structures using a set of features learned from training examples. Lord et al. (2007) demonstrated a groupwise smoothing, segmentation and registration method for cross-sectional MR scans. They developed a level-set framework in which the evolving contours are spatially constrained by an image deﬁned on a common domain, obtained from the ensemble via diffeomorphisms.