Resource atlases for multi-atlas brain segmentations with multiple ontology levels based on T1-weighted MRI

Highlights

• We introduce our T1 multi-atlas inventories (n = 90) covering ages 4–82 years.

• The atlases are defined with hierarchical ontology and comprehensive online repository.

• This rich resource provides flexibility to pre-select atlases for various studies.

• Atlas pre-selection using dynamic age-matching improves segmentation accuracy.

Abstract

Technologies for multi-atlas brain segmentation of T1-weighted MRI images have rapidly progressed in recent years, with highly promising results. This approach, however, relies on a large number of atlases with accurate and consistent structural identifications. Here, we introduce our atlas inventories (n = 90), which cover ages 4–82 years with unique hierarchical structural definitions (286 structures at the finest level). This multi-atlas library resource provides the flexibility to choose appropriate atlases for various studies with different age ranges and structure-definition criteria. In this paper, we describe the details of the atlas resources and demonstrate the improved accuracy achievable with a dynamic age-matching approach, in which atlases that most closely match the subject's age are dynamically selected. The advanced atlas creation strategy, together with atlas pre-selection principles, is expected to support the further development of multi-atlas image segmentation.

Introduction

Quantitative analysis of brain MRI data has played a pivotal role in many brain anatomical studies of development, aging, and various pathological conditions. For image-based quantification, the first and the most important step is to define corresponding brain locations across all participants of the study. One of the most common approaches, which is considered the gold standard in neuroanatomy, is the manual delineation of regions of interest (ROIs). However, because it is labor intensive, manual delineation is usually used for hypothesis-driven studies, in which a small number of target structures are preselected based on a hypothesis. Voxel-based analysis (VBA) is a widely used alternative approach, in which every single voxel is considered an ROI, and corresponding voxel locations are identified across all participants automatically using an image normalization method (Ashburner, 2009, Worsley et al., 1999). Whole-brain structural segmentation is an alternative approach, in which the voxels are joined based on a priori anatomical knowledge, e.g., voxels that belong to the caudate should be joined as one structure (Faria et al., 2010, Fischl et al., 2002, Heckemann et al., 2006, Joshi et al., 2004, Mori et al., 2005, Pham and Prince, 1999, Tu et al., 2008, Tustison et al., 2014, Tzourio-Mazoyer et al., 2002, Woolrich et al., 2009). If the entire brain is segmented into multiple structures, both VBA and segmentation-based methods provide the information about the anatomical features of the entire brain, but from very different granularity levels; in the segmentation-based approach, the information from more than 10⁶ voxels in VBA is greatly reduced to the order of 10².

The meaning of “atlas” varies depending on research fields and study purposes and, thus, clarification is needed. The typical brain atlas consists of images (e.g. histology or MRI) and point-and-annotate labels, describing the locations and names of brain structures. For VBA, atlases mean MR images in specific orientations, positions, and matrix coordinates. They usually do not contain structural labels but the x, y, and z coordinates carry the common anatomical meaning. The images could be chosen within a study or external data such as those from International Consortium of Brain Mapping (ICBM). The MNI coordinates are one of the most widely used standard coordinates and the image could be single-subject such as MNI-Colin27 (Collins et al., 1998) or population-averaged such as linear and nonlinear MNI-ICBM152 (Fonov et al., 2011, Mazziotta et al., 2001). For the segmentation-based analysis (and throughout this paper), the atlas means MRI images with structural labels with three-dimensional boundaries, which carry anatomical references to define structures of interest. One of the simplest approaches to accomplish automated brain segmentation is to warp a single-subject atlas to each subject and transfer the structure labels (Fischl et al., 2002).

In recent years, a multiple-atlas approach has gained popularity due to its superior segmentation accuracy (Artaechevarria et al., 2009, Jia et al., 2012, Langerak et al., 2010, Lotjonen et al., 2010b, Sabuncu et al., 2010, van Rikxoort et al., 2010, Wang et al., 2012, Warfield et al., 2004). In this approach, rather than one reference brain atlas, multiple atlases with consistent structural segmentation are prepared, warped to a subject image, and the multiple segmentation results are fused to achieve the best estimation of the structural identification. Numerous publications have reported improved accuracy using this approach. Most of these previous papers have focused on the algorithms used to fuse the multiple segmentation results. The multiple-atlas approach, however, depends on the availability of atlases with accurate and consistent structural definitions, which is not only labor-intensive, but also has several important issues to be addressed.

For example, one of the most frequently asked questions is how many atlases are needed to improve the segmentation accuracy. Aljabar et al. (Aljabar et al., 2009) has shown that the segmentation accuracy reaches maximum with 15–25 atlases depending on the target structures. However, probably, there is no general answer to this question, because it depends on the anatomical variability among the atlases and the anatomical features of the subject. In an extreme case, if the anatomy of the subject is an outlier, with respect to the anatomical ranges covered by the multiple atlases, we cannot expect high accuracy regardless of the number of atlases used. This naturally extends to the notion of atlas pre-selection based on anatomical or non-anatomical features, such as the ventricle size or age-matching. Another important issue is the way in which anatomy is defined. The minimum definable units are determined by the available anatomical features and contrasts; it is difficult to sub-divide, for example, the globus pallidus (GP) into the internal and external portions, unless MRI provides contrasts to define these intra-GP structures. For segmentation-based analysis, the finer definitions may not always be better. For example, if the entire temporal lobe has 10% atrophy, a report of the volume loss of a dozen of sub-structures within the temporal lobe could not only hurt the statistical power but also result in misleading findings about the regional specificity of the abnormality if the atrophy was statistically detected only in a subset of the constituents. Thus, it is important to test regional specificity by examining larger super structures, such as lobes and hemispheres.

The purpose of this study is to establish a shared atlas resource to support multi-atlas segmentation algorithms for quantitative analyses of brain MR images and test the accuracy levels as well as the efficacy of atlas pre-selection approaches such as dynamic age-matching, which is enabled by the availability of the wide range of the age coverage. In this paper, we introduce our atlas inventories (n = 90) with unique hierarchical structural definitions, which cover ages 4–85 years. The atlas set can be combined with various available multi-atlas fusion algorithms. In addition, we tested the impact of age-matching on the accuracy of the segmentation.