Automated ventricular mapping with multi-atlas fluid image alignment reveals genetic effects in Alzheimer's disease

Abstract

We developed and validated a new method to create automated 3D parametric surface models of the lateral ventricles in brain MRI scans, providing an efficient approach to monitor degenerative disease in clinical studies and drug trials. First, we used a set of parameterized surfaces to represent the ventricles in four subjects' manually labeled brain MRI scans (atlases). We fluidly registered each atlas and mesh model to MRIs from 17 Alzheimer's disease (AD) patients and 13 age- and gender-matched healthy elderly control subjects, and 18 asymptomatic ApoE4-carriers and 18 age- and gender-matched non-carriers. We examined genotyped healthy subjects with the goal of detecting subtle effects of a gene that confers heightened risk for Alzheimer's disease. We averaged the meshes extracted for each 3D MR data set, and combined the automated segmentations with a radial mapping approach to localize ventricular shape differences in patients. Validation experiments comparing automated and expert manual segmentations showed that (1) the Hausdorff labeling error rapidly decreased, and (2) the power to detect disease- and gene-related alterations improved, as the number of atlases, N, was increased from 1 to 9. In surface-based statistical maps, we detected more widespread and intense anatomical deficits as we increased the number of atlases. We formulated a statistical stopping criterion to determine the optimal number of atlases to use. Healthy ApoE4-carriers and those with AD showed local ventricular abnormalities. This high-throughput method for morphometric studies further motivates the combination of genetic and neuroimaging strategies in predicting AD progression and treatment response.

Introduction

The ventricular system is a CSF-filled structure in the center of the brain, surrounded by gray and white matter structures whose volume and integrity are affected by neurodegenerative diseases (Powell et al., 1991). Ventricular changes reflect atrophy in surrounding structures, so ventricular measures and surface-based maps provide sensitive assessments of tissue reduction that correlate with cognitive deterioration in illnesses such as Alzheimer's disease (AD) (Thompson et al., 2004, Carmichael et al., 2006), HIV/AIDS (Thompson et al., 2006) and schizophrenia (Styner et al., 2005, Narr et al., 2001, Narr et al., 2004), and offer a potential approach to evaluate disease progression in large-scale drug trials. In particular, published studies over the past 20 years have found that brain ventricle volume is higher in AD patients than in age-matched healthy controls, and ventricular volume has been proposed as a useful biomarker of disease progression (Thal et al., 2006, Carmichael et al., 2007a, Carmichael et al., 2007b).

Despite years of efforts in automated surface parameterization, the concave shape, complex branching topology (Wang et al., 2005, Ferrarini et al., 2006) and extreme narrowness of the inferior and posterior horns have made it difficult for voxel-based classifiers to create accurate models of the ventricles. It is also difficult for surface parameterization approaches to impose a grid on the entire structure (Wang et al., 2005). In the engineering literature, automated ventricular segmentations often appear with the posterior or inferior horns omitted (Zhu and Jiang, 2004). This is a serious problem as the inferior horns are sensitive to the early degenerative changes in AD. Robust shape models are required to gauge these changes. For the vast 3D MR data sets now being collected (e.g., 339 subjects in Carmichael et al., 2006 and over 3000 scans in the Alzheimer's Disease Neuroimaging Initiative; http://www.loni.ucla.edu/ADNI/), manual segmentations would be impractical due to the time and expertise required; more automated methods are urgently needed.

One promising segmentation approach nonlinearly registers each brain MR volume to a target image volume in which the structures of interest have been labeled and parameterized; the labels are then projected onto the unlabeled data using the computed geometric transformations (Collins et al., 1994, Dawant et al., 1999, Fischl et al., 2002, Svarer et al., 2005, Heckemann et al., 2006). However, the chosen reference image volume influences the results of label propagation; different starting atlases create different segmentations for each subject. In Rohlfing et al., 2003, Rohlfing et al., 2004 and Klein et al. (2005) multiple segmentations were combined, bringing the result closer to manual labels than if only one was used. “Targetless” image warping also combines multiple pairwise or groupwise registrations, reducing registration errors in groups of image volumes (Kochunov et al., 2005, Cootes et al., 2005, Zöllei et al., 2005, Babalola et al., 2006). A related effort in anatomical template construction uses multiple image registrations to create a mean anatomic template that is least distant from a group of anatomies in a strictly defined mathematical sense. Such mean templates may be defined using deformation vector field averaging (Collins et al., 1994, Christensen et al., 2006), by minimizing the mean strain tensors of deformation mappings to the template (Leporé et al., 2007), or by minimizing the total geodesic length of the deformations aligning each individual image volume to the template in the group of diffeomorphisms (Miller, 2004, Lorenzen et al., 2006). Each of these methods has been successful, but has limitations. Deformable atlases that rely on a single atlas template are limited by the unavoidable bias in selecting a specific individual scan as a template, with its unique contrast and geometry. This can make the results depend on the template used (as we show in this paper). Other methods have combined image volumes using multiple registrations. Although this minimizes bias, the resulting template commonly does not provide surface-based anatomical segmentations for the individuals mapped to it, or if it does, it suffers the limitation that a single registration deforms the segmentation in the atlas onto the individual. Surface-based segmentations may also provide additional benefit relative to voxel-based methods that examine image volumes and 3D deformations, as graphical models and maps, such as measures of radial thickness, may be derived from the surface geometry.

In this study, we validated a new automated technique for 3D segmentation that combines multiple registration-based surface models into one. We show that this increases the label propagation accuracy and the power to detect disease effects, without requiring any interactive human input other than the initial expert labeling of a small set of image volumes. Combining segmentations is not a new idea, but the approach has not been previously applied and validated for surface-based analyses of anatomy.

We first linearly registered MRI scans to a standard space. A small subgroup of image volumes was randomly chosen and manually traced. Lateral ventricular models were created in these image volumes and converted into parametric surface meshes (we will call these labeled image volumes “atlases”). We then fluidly registered each of these atlases and mesh models to all other subjects. A mesh averaging technique then combined the resulting fluidly propagated surface meshes for each image volume. To optimize the algorithm, we examined how the segmentation error (assessed using the symmetrized Hausdorff distance and 2-norm) and the disease detection power (assessed using cumulative distributions of surface statistics and the False Discovery Rate) depended on the number of atlases, leading to an empirical basis for choosing the number of templates in future studies that use multi-atlas segmentation.

We used this method to analyze morphometric differences in the lateral ventricles of subjects with AD, versus healthy elderly controls. The lateral ventricles of AD patients have been widely studied, and are known to display increased atrophy rates in surrounding tissues compared to aged matched controls, and are therefore an ideal system in which to validate our method.

Carriers of the apolipoprotein E epsilon-4 gene (ApoE4) are at heightened risk for early development of AD (Corder et al., 1993, Strittmatter et al., 1993). Structural imaging studies of AD patients have revealed morphological deficits in carriers of the ApoE4 allele versus non-carriers (Lehtovirta et al., 1995, Juottonen et al., 1998, Geroldi et al., 1999, Geroldi et al., 2000, Mori et al., 2002). In healthy elderly subjects, the ApoE4 allele is associated with diminished CNS glucose utilization (Reiman et al., 2004) and some studies have reported impairment in cognitive function (Deary et al., 2002). Inheritance of the ApoE4 allele is a risk factor for early development of AD, and, in epidemiological studies of several populations, between 10% and 18% of elderly healthy controls and 24–52% of AD patients have been found to carry at least one copy of the ApoE4 allele (Beyer et al., 1997). ApoE4-carriers with MCI (mild cognitive impairment) may respond especially well to cholinesterase treatment. In a study of 768 subjects with amnestic MCI, the ApoE4 carriers treated with donepezil showed a significant reduction in progression to AD at 3 years compared to carriers who received placebo, but the non-carriers did not (Petersen et al., 2005). There is, therefore, some interest in establishing whether neurodegeneration progresses differently in ApoE4-carriers than non-carriers and how atrophic changes interact with genotype. It is not yet known whether the ventricles of ApoE4 carriers differ from those of non-ApoE4 carriers. We thus validated our method further by looking at ventricular shape differences in a case that has not yet been studied, but where such changes are expected to exist.