Iterative multi-atlas-based multi-image segmentation with tree-based registration

Abstract

In this paper, we present a multi-atlas-based framework for accurate, consistent and simultaneous segmentation of a group of target images. Multi-atlas-based segmentation algorithms consider concurrently complementary information from multiple atlases to produce optimal segmentation outcomes. However, the accuracy of these algorithms relies heavily on the precise alignment of the atlases with the target image. In particular, the commonly used pairwise registration may result in inaccurate alignment especially between images with large shape differences. Additionally, when segmenting a group of target images, most current methods consider these images independently with disregard of their correlation, thus resulting in inconsistent segmentations of the same structures across different target images. We propose two novel strategies to address these limitations: 1) a novel tree-based groupwise registration method for concurrent alignment of both the atlases and the target images, and 2) an iterative groupwise segmentation method for simultaneous consideration of segmentation information propagated from all available images, including the atlases and other newly segmented target images. Evaluation based on various datasets indicates that the proposed multi-atlas-based multi-image segmentation (MABMIS) framework yields substantial improvements in terms of consistency and accuracy over methods that do not consider the group of target images holistically.

Fig. 1

The overall framework of MABMIS. The three key components are represented by the rectangular boxes.

Fig. 2

Two intermediate templates guided registration schemes: intermediate templates generation (ITG, left) and intermediate templates selection (ITS, right). In ITG, one image is generated, by warping the template with a known deformation field, as the intermediate template for each test image. This intermediate template is more similar to the test image than the original template; therefore, the residual deformation field between each test image and the intermediate template can be estimated with higher accuracy. In ITS, on the other hand, all test images are organized into a tree structure with the template as the root. The deformation field between each test image and the template is thus the combination of several smaller ones, thus joining several intermediate templates along the path traversed by the test image to reach the template.

Fig. 3

Flowchart illustrating the registration based on the combinative tree.

Fig. 4

Illustration of a combinative tree, which can be expanded by appending the target images. For each target image, its best-matching image on the current tree can be an original atlas, a simulated atlas, or even another target image as shown within the green circle. Some images (e.g., the left-most atlas) may be connected to the root template via more than one path.

Fig. 5

Two different ways of aligning the atlases to the target image.

Fig. 6

Sequential registration and segmentation of the target images.

Fig. 7

Sample images from the ADNI dataset. Large anatomical differences remain even after affine registration: MCI patients (top row) and normal controls (bottom row).

Fig. 8

The distribution of intensity differences of all atlases after registration by four different methods on ADNI dataset.

Fig. 9

The proposed method (MABMIS) achieves the best overlap rates for 46 out of 54 ROIs.

Fig. 10

Comparison of segmentation accuracy. From left to right: (1) the label propagation result using a best-matching atlas, (2) the label fusion results with the pairwise registration, the statistical model based registration, the tree-based registration, and MABMIS with iterative updating, and (3) manual labels (ground-truth).

Fig. 11

Segmentation consistency measured by the average overlap rates given by different multi-atlas-based image segmentation methods.