MORPHOMETRIC ANALYSIS OF HIPPOCAMPUS AND LATERAL VENTRICLE REVEALS REGIONAL DIFFERENCE BETWEEN COGNITIVELY STABLE AND DECLINING PERSONS

Abstract

Alzheimers disease (AD) is a progressive neurodegenerative disease most prevalent in the elderly. Distinguishing disease-related memory decline from normal age-related memory decline has been clinically difficult due to the subtlety of cognitive change during the preclinical stage of AD. In contrast, sensitive biomarkers derived from in vivo neuroimaging data could improve the early identification of AD. In this study, we employed a morphometric analysis in the hippocampus and lateral ventricle. A novel group-wise template-based segmentation algorithm was developed for ventricular segmentation. Further, surface multivariate tensor-based morphometry and radial distance on each surface point were computed. Using Hotellings T² test, we found significant morphometric differences in both hippocampus and lateral ventricle between stable and clinically declining subjects. The left hemisphere was more severely affected than the right during this early disease stage. Hippocampal and ventricular morphometry has significant potential as an imaging biomarker for onset prediction and early diagnosis of AD.

1. INTRODUCTION

As one of the most common neurodegenerative diseases in older people, Alzheimers disease (AD) affects cognitive function presumably by accumulation of amyloid-β (Aβ) and hyperphosphorylated tau proteins in the brain. The increasing number of aging patients suffering with AD speaks strongly to the importance of developing effective methods for early diagnosis. Amnestic Mild Cognitive Impairment (aMCI), with subjective and objective evidence of memory loss while generally maintaining independence of function in daily life [1], is regarded as a precursor of AD. However, evidence suggest that disease-related neurobiological changes have already taken place prior to the onset of the overt symptoms [2]. It is commonly agreed that an effective presymptomatic diagnosis and intervention of AD could have enormous public health benefits. Brain imaging has the potential to provide valid diagnostic biomarkers of AD risk factors and preclinical stage AD. MRI-based volumetric measures in several structures (as reviewed in [3]), including whole-brain, entorhinal cortex, hippocampus, and temporal lobe volumes, as well as ventricular enlargement, correlate closely with changes in cognitive performance, supporting their validity as markers of disease progression. To date, few studies have investigated any sub-cortical and lateral ventricular morphometry that are associated with future cognitive decline in an asymptomatic normal aging population.

Although most subcortical morphometrical studies used volume as the atrophy measurement [4, 5], recent researchers [6, 7] has demonstrated that subcortical structure surfaces may offer advantages over volumetric measures. Surface-based methods study patterns of subfield atrophy or enlargement and may identify detailed point-wise correlation between atrophy and cognitive functions/biological markers. Statistics derived from anatomical surface models, such as radial distances (RD, distances from the medial core to each surface point) [8, 6], spherical harmonic analysis [9, 10], local area differences (related to the determinant of the Jacobian matrix) [11], and Gaussian random fields [12] have all been applied to analyze subcortical morphometry. Surface tensor-based morphometry (TBM) [13] is an intrinsic surface statistic that examines spatial derivatives of the deformation maps that register brains to a common template and construct morphological tensor maps. Recently, we developed multivariate TBM (mTBM) [14] and it showed good sensitivity in detecting group differences on surface models [14]. To get a complete set of surface statistics, we further combined the mTBM together with RD as those two statistics were complementary [15]. In this work, we evaluated the potential of surface mTBM and RD as imaging biomarkers for early indications of future cognitive decline in a healthy control (CTL) population.

In this study, we first developed a group-wise template-based lateral ventricle segmentation algorithm. This method can capture the group variations by creating a group-wise template with geodesic shooting. Secondly, we applied our brain surface morphometry pipeline to analyze brain MRIs obtained from the well-characterized Arizona APOE cohort of presymptomatic individuals. We hypothesized that our surface multivariate statistics may identify subtle shape difference on both hippocampus and lateral ventricle on presymptomatic subjects. The results showed that our new method identified significant regional differences between stable and declining participants. Our results may help improve the sensitivity and accuracy of hippocampal and ventricular morphometry for preclinical AD research.

2. METHODS

2.1. Subjects and Data Acquisition

Since 1994, cognitively normal residents of Maricopa County age 21 years and older were recruited through local media ads into the Arizona APOE cohort [16]. Demographic, family, and medical history data were obtained on each individual undergoing APOE genotyping. All individuals gave their written, informed consent, approved by the Institutional Review Boards of all participating institutions. Genetic determination of APOE allelic status was performed using a polymerase chain reaction (PCR) based assay. Screening tests included a medical history, neurologic examination, the Folstein Mini-Mental Status Exam (MMSE), the Hamilton Depression Rating Scale (Ham-D), the Functional Activities Questionnaire (FAQ), Instrumental Activities of Daily Living (IADL), and Structured Psychiatric Interview for DSM-IIIR. There were no potentially confounding medical, neurological, or psychiatric problems. None met the published criteria for MCI, AD, or any other form of dementia, or major depressive disorder at the time of their enrollments. A subset of 18 subjects with presymptomatic imaging studies and who have converted to aMCI or AD 2 years, on average, after their last imaging visits was selected as the declining group. Another group of 35 subjects matched for sex and age and who remained cognitively unimpaired for at least 4 years were selected as the comparison group, stable control group. Their baseline high-resolution T1 MRI scans were used in this study.

2.2. Segmentation of Hippocampus and Lateral Ventricle

The hippocampus was segmented from T1 images by using FSL software package [17]. Details of processing steps were reported in our prior work [18]. For ventricular segmentation, in this research, we proposed a novel pipeline that computed a group-wised ventricular template and used it to accurately segment continuous ventricular structures.

Basically, the T1 structural images were linearly registered into a standard space (MNI152) before segmentation. Then the registered images were parcellated into three brain tissue types (the gray matter, white matter and CSF) using a modified Guassian mixture model (SPM8 packages, http://www.fil.ion.ucl.ac.uk/spm/) to estimated the probability of the tissue type each voxel belongs to based on a prior probability map. After doing tissue segmentation for all subjects, we created a group-wise CSF template by applying the geodesic shooting algorithm [19] to all individual CSF masks. Essentially, in the first step, the initial CSF template was set with mean shape and intensity of the subjects’ CSF binary mask in the form of a probability map and was updated iteratively. During each iteration, the objective function,

$$E=\frac{1}{2}{\Vert L{\upsilon }_0\Vert }^2+\frac{1}{2{\sigma }^2}{\int }_{x\in \omega }|{\mathbf{J}}_1^{\phi }(x)|{(f({\phi }_1(x))-\mu (x))}^{2}dx$$ (1)

was minimized by a proper initial velocity field υ₀. The cost function dynamically validated the accuracy of the diffeomorphic mapping. Here, φ₁ is a forward diffeomorphism from subject to template guided by υ₀. L is a linear differential operator:

$${\Vert L{\upsilon }_0\Vert }^2={\int }_{x\in \omega }(\frac{{\lambda }_1}{4}{\Vert D\upsilon +{(D\upsilon )}^T\Vert }^2+{\lambda }_{2}tr{(D\upsilon )}^2+{\lambda }_3{\Vert \upsilon \Vert }^2)dx$$ (2)

Eventually, the stable result of CSF template is regarded as the group-wise CSF template. As a part of the CSF, the group-wise ventricular template were subsequently obtained by mapping a probalbility ventricular mask onto the group-wise CSF template. In this study, the ALVIN (Automatic Lateral Ventricle delIneatioN [20]) binary mask was applied to exclude CSF tissues outside the lateral ventricle. The whole ventricular segmentation pipeline is illustrated in Fig. 1.

Fig. 1.

Ventricle Segmentation Pipeline. The order of processing is followed by A-B-C-D.

2.3. Surface Multivariate Tensor-based Morphometry

Based on segmented binary volume masks, we extracted hippocampal and ventricular boundaries with a topology preserving level-set method [21] and constructed triangular surface meshes with marching cubes algorithm [22]. Later, the surfaces were further smoothed using a two step mesh smoothing method, “progressive meshes” and Loop subdivision, which has been proved to be feature-preserving meanwhile effectively reduce the noise and partial volume effect [23]. Fig. 2 shows some reconstructed surfaces.

Fig. 2.

Segmentation Results and Surface Reconstruction. We overlaid the segmented ventricular and hippocampal volume masks on the original T1 image and showed the surface reconstruction meshes based on the volume masks. L stands for the left side and R means the right side.

For surface registration, we need to compute the conformal grids onto the surfaces. As hippocampus and ventricle have their specific topology, we treat them differently. Specifically, a hippocampal surface may be presented as a genus zero surface with two open boundaries linked two holes, one hole locates at the front while another at the back (we called it as topology optimization [15]). This enabled us to unfold a hippocampal surface (similar to a cylinder) to a rectangle. However, a ventricular surface is hard to follow the foregoing processing because of its “multiple-arm” structural property. Therefore, we parcellate the whole ventricular surface into three sub-structures by the holomorphic flow segmentation method [14]. Three horns, including anterior horn, posterior horn and inferior horn, were automatically located and separated from each ventricular surface. The horn-division principle is followed the topological properties of the lateral ventricle thus brings less errors to later group analysis.

Next, constrained harmonic maps were applied to respectively register the hippocampal and separated ventricular surfaces to the standard maps. The harmonic mapping τ can be expressed as below:

$$\tau ◦{\tau }_1(S_1)={\tau }_2(S_2),\tau ◦{\tau }_1(\partial S_1)={\tau }_2(\partial S_2),\delta \tau =0.$$ (3)

Here τ₁ and τ₂ are conformal parameterizations that respectively maps surface S₁ and S₂ to R². The map ϕ from surface S₁ directly to surface S₂ can be obtained by $\varphi ={\tau }_1◦\tau ◦{\tau }_2^{-1}$.

After surface registration, we calculated the vertex-wise surface multivariate statistics, which can be represented as a 4 × 1 feature vector consisting of two parts. The first part is surface multivariate tensor-based morphometry (mTBM), a “Log-Euclidean metric” [24] on the set of deformation tensors S, i.e., a 3 × 1 positive definite matrix (log(S)). The deformation tensors S is computed as S = (J^T J)^1/2, where J is the Jacobian matrix [14]. Suppose there are two faces A = [a₁, a₂, a₃] and B = [b₁, b₂, b₃], which were isometrically embedded on the Euclidean space. The discrete derivative map J from A to B is J = [a₃−a₁, a₂−a₁][b₃−b₁, b₂−b₁]⁻¹. The second part of multivariate statistics is the radial distance [8, 25], a 1 × 1 vector, which measures the shortest distance from each surface point to the medial axis of a tube-shape surface.

2.4. Group Difference Study

The Hotellling’s T² test was performed to evaluate the morphometric variations of hippocampal and ventricular surfaces between two groups of subjects, declining controls vs stable control group on each vertex. Statistical results were corrected for multiple comparison using the permutation test [14]. Basically, we calculated the Mahalanobis distance based on the true group labels first. Then we randomly assigned the object surfaces into two groups which had the same number of subjects as in the true group and re-computed the group distance on each surface point. This process repeated 5, 000 times with the outcome of 5, 000 permutation values on each vertex. A probability (uncorrected p value) on each surface point was computed as the ratio of the number of permutation values greater than the true group t value to the total permutation times. This result was shown in the form of a p-map of the test. After that, given a pre-defined statistical threshold, e.g. p < 0.05, we defined a feature that to be the number of surface points with uncorrected p value lower than this threshold. The feature could be regarded as the real effect in the true experiment and by comparing it to the features derived from the random groupings, we obtained a ratio stood for the fraction of the time an effect of similar or greater magnitude to the real effect occurs in the random assignments. This ratio, the overall (corrected) significance, provided a global significance level of the map.

3. EXPERIMENTAL RESULTS

3.1. Hippocampal and Ventricular Surface Reconstruction

The hippocampal surface reconstructive procedure was reported in our prior work [18] which has also been applied to surface-based morphometric analysis. For ventricular segmentation, the group-wise ventricular probability template was created using geodesic shooting. We chose p = 10% as the threshold value to create a binary volume mask and consequently warped it back to each individual space according to the inverse matrix. Both hippocampal and lateral ventricular masks in the individual space were mapped onto the original T1 images for quality inspection of segmentation. Here, we compared our proposed ventricle segmentation with other previous methods, e.g. ALVIN [20] and FreeSurfer [26], as well as the manual segmentation assessed by Dice coefficient [27] (ranging from 0 to 1) where higher value stands for higher similarity between segmented results of our method and others (Table 1). It is worth noting that our segmentation method could reconstructed a whole connected lateral ventricular mask which benefits the surface-based morphometric analysis while preserving accurate shape information. After converting binary masks to mesh surfaces, a feature-preserved smooth method worked on each of the surfaces. Fig. 2 illustrates the surface reconstruction procedure and the overlaid results.

Table 1.

Validation of the segmentation results

Method	Dice coefficient
	Left LV	Right LV
ALVIN	0.91	0.91
Freesurfer	0.86	0.85
Manual	0.90	0.89

3.2. Morphometric Difference between Stable and Declining Controls

We computed the surface morphometric difference of hippocampus and lateral ventricle between stable and declining individuals for the purpose of evaluating the affection of cognitive decline to hippocampus and ventricle. All subjects were cognitively unimpaired when the baseline images were acquired and later on, some of them developed clinically significant cognitive decline while the rest still remained unimpaired. We used the Mahalanobis distance to measure the difference between the mean morphometry feature vectors between these two groups, i.e. stable and declining subjects. P-maps computed by Hotelling’s T² test were shown in Fig. 3. The statistical results were corrected for multiple comparisons with permutation test. With p < 0.05 as the threshold, the global significance levels on hippocampus and lateral ventricle reached the significant level on both sides (hippocampus, left p = 0.0135, right p = 0.0367; ventricle, left p = 0.0131, right p = 0.0461).

Fig. 3.

Statistical Result of Group Difference on Hippocampus and Lateral Ventricle. The middle color bar indicated the range of p values. L stands for the left side while R means the right side.

In particular, we identified that strong difference on the left sides of both hippocampus and lateral ventricle. It also verified some of our prior work on ApoE research [18]. It may suggest that the cognitive decline related to brain atrophy started from left side and subsequently extend to the right.

Overall, the experimental results showed a strong indication that our system may be able to capture the subtle difference of abnormal degrees of hippocampal atrophy and ventricular enlargement that are associated with future symptomatic memory loss. Our work may be beneficial for precise analysis of MRI patterns for the preclinical AD research.

4. CONCLUSION AND FUTURE WORK

In the first part of our work, we proposed a novel ventricular segmentation algorithm. There are a few advantages of this automatic segmentation pipeline. First, the individual ventricle mask is derived from a group template which reflects the group variation. Second, it generates a whole connected 3D shape model which benefits the surface-based morphometric analysis. Last but not least, it is an automatic pipeline without much subjective intervention during the process. In the second part of our research, we studied the morphometry using surface mTBM [14, 15] on hippocampus and lateral ventricle between stable and declining subjects. Our work showed that hippocampal and ventricular morphometry may be useful as imaging biomarkers to predict imminent cognitive decline in asymptomatic, cognitively unimpaired subjects. Our future work will focus on developing advanced machine learning algorithms which make use of our surface mTBM features to advance the preclinical AD research.