Structural Changes in Thalamic Nuclei Across Prodromal and Clinical Alzheimer’s Disease

INTRODUCTION

Alzheimer’s disease (AD), the most prevalent form of dementia afflicting over 5.8 million people in the United States [1], has long been linked to pathological changes seen in the medial temporal lobe, mainly the hippocampus [2, 3]. The hippocampus has a well-characterized role in episodic memory [4], the decline of which is a hallmark of AD. Not surprisingly, neuroimaging studies in AD invariably show substantial atrophy of the hippocampus [5–10]. While the hippocampus certainly plays a significant role in the pathogenesis of AD, there is growing evidence that the hippocampus is part of a larger network of brain regions implicated in episodic memory, namely the limbic memory circuit, or the Papez circuit [11]. The limbic memory circuit is a set of cortical and subcortical structures and their interconnections that includes the hippocampus, anterior thalamic nuclei (anteroventral, anterodorsal, anteromedial), the fornix, mammillary bodies, and the posterior cingulate region [11]. Given their established role in episodic memory [12], changes anywhere within this circuit could play a role in the memory loss associated with AD. In addition to the limbic memory circuit, the mediodorsal (MD) nucleus has also been shown to serve an important role in memory in conjunction with the perirhinal cortex [13]. However, despite their established role in memory, the anterior nuclei of the thalamus have received very little attention in AD research. Braak et al. [14] found neurofibrillary tangles and amyloid plaques within the anterior nuclei of the thalamus, with the anterodorsal nucleus being the most affected. These findings were confirmed by Rub et al. [15] who found early neurofibrillary tangles in the laterodorsal and anterodorsal nuclei of the thalamus. Further, Ryan et al. [16] found atrophy in thalamic and caudate volume in presymptomatic familial AD even prior to atrophy in hippocampal volume.

While a handful of imaging studies have reported significant decreases in whole thalamus volumes in mild cognitive impairment (MCI), cognitive aging, and AD [5, 17–19] and another has reported regional changes in shape of the dorsomedial thalamus in AD [20], very few studies have focused on individual thalamic nuclei. This is, in large part, due to technical challenges in successful and accurate parcellation of the thalamic nuclei. Conventional cerebrospinal fluid-nulled (CSFn) Magnetization Prepared Rapid Gradient Echo (MPRAGE), which is T₁-weighted, and T₂-weighted magnetic resonance imaging (MRI) pulse sequences have poor intra-thalamic nuclear contrast. Most attempts to parcellate the thalamus have been based on diffusion MRI techniques [21–25], but this modality is limited by low spatial resolution and a lack of significant diffusion anisotropy in the largely gray-matter thalamus, resulting in poor delineation of small structures such as anteroventral or geniculate nuclei. Others have used diffusion MRI tractography to identify thalamic nuclei based on cortical connections [26–28]. However, the delineated regions tend to be large, and are not based on inherent tissue differences within the thalamic nuclei. More recently, structural imaging-based techniques for parcellating the thalamus have emerged. In one technique, manual segmentation on a set of histological and ex vivo imaging data are combined to create an atlas, which was then used to segment the thalamus in vivo data using Bayesian inference [29]. Another technique used a multi-atlas approach to segment the thalamus based on a hierarchical statistical shape model [30]. Recent work by Su et al. [31] has demonstrated that a specialized white-matter-nulled (WMn) MPRAGE sequence produces increased contrast within the thalamus. This was combined with a multi-atlas joint label fusion technique to produce parcellations of the thalamus into distinct nuclei in a technique called THalamus Optimized Multi Atlas Segmentation (THOMAS).

In the very few studies that have investigated differences in thalamic nuclei in MCI and AD, there have been varied results. Iglesias et al. [29] found that in addition to the whole thalamus, six thalamic nuclei, including the anteroventral nucleus, the mediodorsal nucleus, and medial geniculate nucleus, were significantly smaller in a large cohort of only AD subjects (Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort) when compared to healthy controls. In another study, Low et al. [32] found statistically significant differences in volumes among the anterior, lateral, and posterior clusters of thalamic nuclei between healthy controls and AD subjects. However, they found no differences in the volume of the whole thalamus among healthy controls and subjects with MCI and AD despite using the same technique as Iglesias et al. The major finding of their work is that while there is no difference in absolute volumes of the ventral group of thalamic nuclei among the three groups, the ventral nuclei are significantly smaller within the left thalamus than the right thalamus in subjects with AD (i.e., left-right asymmetry), a phenomenon that has been reported for many other brain regions in AD as well [7, 33–37].

In this work, we use a novel thalamic segmentation technique (THOMAS) to investigate changes in thalamic nuclear volumes in subjects with increasingly severe cognitive impairment from healthy controls to AD. We take advantage of the large ADNI database, which includes high-quality structural MRIs for healthy controls, biomarker confirmed subjects with early and late MCI, and AD to get insights into changes in thalamic nuclear volumes at different stages of the disease. With the increasing evidence of thalamic involvement in the progression of AD, we expected to find increasing atrophy of the AV and MD nuclei of the thalamus throughout the stages of increasing cognitive impairment up to AD. Specifically, we predicted that atrophy of these select nuclei will correlate with clinical and neuropsychological measures of memory function and cognition.

METHODS

Data used in preparation of this article were obtained from the ADNI database (http://adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, positron emission tomography, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD. For up-to-date information, see http://www.adni-info.org.

Data processing

Thalamic segmentation was implemented using a modification of the original THOMAS method of Su et al. [31] briefly described below. The eleven delineated nuclei are grouped as follows:

Medial group: mediodorsal (MD) nucleus, centromedian (CM) nucleus, habenula (Hb).

Posterior group: Pulvinar (Pul) nucleus, medial geniculate nucleus (MGN), lateral geniculate nucleus (LGN).

Lateral group: Ventral posterolateral (VPL), ventral lateral anterior (VLa) nucleus, ventral lateral posterior (VLp) nucleus, ventral anterior (VA) nucleus.

Anterior group: Anteroventral (AV) nucleus.

In 9 healthy control WMn-MPRAGE datasets, eleven thalamic nuclei and the mammillothalamic tract (MTT) were manually segmented by an expert neuroradiologist using the Morel stereotactic atlas as a guide to create a multi atlas. A mean brain template was created from the library of manually segmented datasets (priors). The buildtemplate feature of Advanced Normalization Tools (ANTs) package [39] was used to iteratively register each prior to an average of the priors and then to create a mean template by averaging the registered priors, which has excellent SNR and image contrast. To segment the thalami of individual subjects, the template image was first registered to the subject’s T1-weighted image using the nonlinear symmetric image normalization (SyN) algorithm implemented in ANTs. Each anatomical prior was also registered to the template image and these were available a priori [39, 40]. A single composite transformation to warp each anatomical prior to each subject’s T1-weighted image was then generated by combining the prior to template warp with the template to subject warp. This composite transformation was applied to all thalamic nuclei labels from each of the anatomical priors, to produce 9 sets of thalamic nuclei labels aligned with each subject’s image. Finally, the 9 sets of labels were fused into a single set of labels using majority voting as implemented in ANTs, producing a single set of thalamic nuclei labels aligned to each individual subject. These steps are shown in Fig. 1. An example T1 MPRAGE image segmented using the modified THOMAS algorithm is shown in three planes in Fig. 2.

OPEN IN VIEWER

Fig. 1

Multi-atlas segmentation scheme for thalamic nuclei segmentation. The multi-atlas consists of 20 manually segmented WMn MPRAGE datasets, which are warped to subject space and label fused using a majority voting scheme. A WMn template is used as an intermediate step to improve robustness and cropping is performed to improve speed and accuracy.

OPEN IN VIEWER

Fig. 2

Thalamic nuclei segmentation labels from the modified THOMAS method overlaid on MPRAGE on a representative healthy control subject (top row) and a representative Alzheimer’s disease subject (bottom row).

These segmented labels were then used to estimate the volume of each thalamic nucleus. In addition to nuclei volumes, a laterality index (LI) was also calculated for all thalamic nuclei as LI = (L-R) / (0.5 ^* (L + R)) ^* 100%as described by Low et al. [32] In addition to thalamic nuclei volumes, the volumes of bilateral hippocampi were estimated using FreeSurfer (version 7.0.0). Intracranial volumes (ICV) were also computed for each subject using FreeSurfer’s recon-all command.

Note that in the original implementation of THOMAS, joint fusion was used to combine the labels as the input images were white-matter-nulled MP-RAGE images. In order to validate the accuracy of the modified THOMAS method using conventional CSFn-MPRAGE versus WMn-MPRAGE, we performed a comparison on 18 healthy subjects, where both sequences were acquired. The WMn-MPRAGE was segmented using THOMAS with label fusion as described in Su et al. and served as a “gold standard”. The CSFn-MPRAGE data was segmented using THOMAS with majority voting as described above. Accuracy of the proposed majority voting-based THOMAS method was assessed by computing Dice coefficients and a volume similarity index (VSI) between the results obtained from the majority voting method technique compared to those obtained from the WMn-MPRAGE data using joint fusion THOMAS algorithm. Dice coefficients and the volume similarity index (VSI) are calculated as: DICE Coefficient = 2 | X ∩ Y | | X | + | Y | and VSI = 1 - abs ( | X | - | Y | ) | X | + | Y |

where X and Y refer to the two segmentation labels being compared, with one being the ground truth (the WMn MPRAGE results in this case). |X| and |Y| refer to the number of voxels in X and Y respectively.

RESULTS

Comparison of thalamic nuclei volumes

Initial ANCOVA analysis using age, ICV, biological sex, and years of education revealed that biological sex and years of education did not significantly impact measures of volume for any of the thalamic nuclei nor the hippocampus, and were thus removed from the model. Thus, only age and ICV were included as covariates for this analysis.

The volumes of individual thalamic nuclei, as well as the whole thalamus and the hippocampus for all subjects across all four study groups are shown in Fig. 3. Post hoc analysis revealed that in the bilateral MD nuclei, the left Pulvinar nucleus, and the left MGN nucleus, there were significant differences in the volumes between healthy controls and in subjects with EMCI. The volumes of the bilateral AV nuclei, bilateral MD-Pf, the bilateral Pulvinar nuclei, the bilateral CM nuclei, the left MGN nucleus, and the entire thalamus bilaterally were significantly smaller in LMCI subjects when compared to HC subjects. Further, the volumes of the AV nucleus, the Pulvinar nucleus, the MGN, the CM nucleus, and the MD-Pf nucleus in subjects with AD were significantly smaller bilaterally compared to HC subjects. The hippocampus was significantly smaller bilaterally in EMCI, LMCI, and AD subjects when compared to the HC group. The full list of statistical results for the above comparisons are detailed in Supplementary Table 2.

OPEN IN VIEWER

Fig. 3

Thalamic nuclei volumes plotted as box and whisker plots compared across four groups. In the above plots, healthy controls (HC) are plotted in blue, early mild cognitive impairment (EMCI) is plotted in green, late mild cognitive impairment (LMCI) is plotted in orange, and Alzheimer’s disease (AD) is plotted in red. Each subject’s nuclei volume is plotted as a filled circle on the plot. The black “x” on each box and whisker plot denotes the mean volume for each group. Statistically significant differences (p < 0.05) in pairwise comparisons after Dunnett’s test are shown as black bars above the box and whisker plots.

Effect size for each of the above post-hoc analyses is plotted in Fig. 4 as the Cohen’s d score. The effect size of HC versus EMCI is plotted in blue, the effect size of HC versus LMCI is plotted in orange, and the effect size of HC versus AD is plotted in yellow. On both the left and the right side of the brain, the whole thalamus has a very large effect size (d = 0.70 on the left, and d = 0.76 on the right), followed by the Pulvinar nucleus (d = 0.68 on the left and d = 0.76 on the right) when comparing HC to AD. As expected, the majority of the nuclei show progressively larger effect sizes with increasing disease severity.

OPEN IN VIEWER

Fig. 4

Effect sizes plotted as Cohen’s d computed from pairwise comparisons of healthy controls (HC) with early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), and Alzheimer’s disease (AD), in the thalamic nuclei shown to have statistically significant differences in volume.

In addition to comparing nuclei volumes, the LI was used to compare asymmetry in the atrophy of nuclei with disease severity. ANCOVA analysis showed no statistically significant differences in the LI (data not shown) for any of the thalamic nuclei. While there was no statistically significant difference in laterality of the nuclei with increasing disease severity, there were several nuclei that were consistently smaller on one side than the other across all disease states. The AV, VA, and MGN nuclei were all generally smaller in the left hemisphere.

DISCUSSION

In this study, we investigated changes in the volumes of thalamic nuclei throughout disease progression, i.e., early MCI, late MCI, and finally AD using a novel, accurate thalamic segmentation method. We showed that there are statistically significant differences in several thalamic nuclei at different stages of cognitive impairment, with increasingly smaller volumes with increasing disease severity. Notably, the AV nucleus, a component of the limbic memory circuit, was significantly smaller in subjects with LMCI and AD than in HC subjects. Further, the MD-Pf nucleus, a structure with a well-known role in memory, was also significantly smaller in subjects with EMCI, LMCI, and AD. Additionally, the CM nucleus, which has been shown to have numerous connections within the limbic system [41], was significantly smaller in LMCI and AD than in HC subjects. The findings presented in this work highlight the utility and sensitivity of the thalamic segmentation algorithm and are consistent with other findings suggesting that pathological changes in the thalamus, and more broadly, the limbic memory circuit, play a significant role in the progression of AD.

Notably, the AV nucleus is demonstrably smaller in subjects with increasingly severe cognitive impairment. This result is consistent with the hypothesis that the anterior thalamus plays an important role in the development of the memory impairment that characterizes the early stages of AD [11, 42]. More broadly, Argyropoulos et al. have shown that consideration of structural/functional changes within the entire Papez circuit better explains declining memory performance in subjects with autoimmune limbic encephalitis than hippocampal atrophy alone [42]. The findings presented here and those described elsewhere suggest that a more comprehensive analysis of the entire limbic memory circuit, including the thalamus, the hippocampus, the cingulate gyrus, and their white matter connections may provide more complete insight into the neural substrates of memory loss in AD. Other nuclei from the limbic system, namely the MD-Pf and the CM nuclei showed similar patterns of decreasing volume with disease severity, further promoting the idea that memory loss in AD involves damage to a network of limbic structures and their connections.

In addition to the limbic nuclei discussed above, the pulvinar nucleus and the MGN also showed significant differences in volume with disease severity. In the work by Iglesias et al. [29], they demonstrated a statistically significant difference in the volume of the MGN in subjects with AD. Further, amyloid plaques have been documented throughout the pulvinar nucleus in patients with AD [43]. The pulvinar nucleus has widespread connections throughout the cortex, including visual areas and memory-related regions within the default mode network such as the lateral and medial parietal cortex, as well as the precuneus and parahippocampal gyri. Disruption of these thalamocortical networks may contribute to visual and memory disturbances in AD [44].

Low et al. [32] demonstrated asymmetrical atrophy of the ventral thalamic nuclei in AD, with the left nuclei atrophying significantly more than the right nuclei. In the present work, we were unable to replicate this finding for any thalamic nuclei. There are several plausible explanations for this discrepancy. In their work, there are relatively few subjects (n = 65) relative to what was used here (n = 540) from the ADNI study, making their analysis more susceptible to Type II errors. Further, they utilized a different thalamic segmentation software implemented in FreeSurfer which was demonstrated to be less reliable (as quantified by poorer dice coefficients using manual segmentation as gold standard) than THOMAS [31]. Other studies have found similar patterns of asymmetrical atrophy for other structures in the brain [37, 45–47], however, so further investigation into the findings by Low are certainly warranted.

While this study took advantage of the large database provided by ADNI, there are a number of limitations that need to be addressed in future investigations. The template and multi-atlas used in this study were based on white-matter nulled T1 images due to the increased contrast of thalamic nuclei using that imaging modality. However, this is not a conventional MRI sequence, and there is limited data available to perform the above analysis. Despite the differing contrast between white-matter nulled and CSF-nulled T₁-weighted images, good to excellent reliability of the registration and volume estimation has been demonstrated here (see Supplementary Material) [47, 48]. While this work included healthy controls, EMCI, LMCI, and AD subjects, there was no analysis of thalamic nuclei volumes as HC subjects progress to AD over time, which would arguably provide even more useful data with respect the progression of the disease. Not all subjects included in this analysis with EMCI will progress to LMCI and AD, and thus may show a different pattern of atrophy than subjects with presymptomatic AD. There were simply not enough imaging studies collected at 3T to perform this type of analysis using the ADNI dataset. Further, the distinction between EMCI and LMCI defined in the ADNI, which distinguishes the two based on a single memory test, has been shown to have relatively high false-positive rates [49], an issue that must be addressed in the future if the progression to AD is to be more accurately followed. If the hypothesis that AD is a result of disease throughout the limbic memory circuit, and not just the medial temporal lobe is true, then we would expect to see changes in the white matter pathways that connect the gray matter structures within this circuit. In future work, it will be worthwhile to include white matter metrics, such as those derived from diffusion MRI experiments, to further characterize network-level changes occurring with disease progression.

Many of the above shortcomings of the present study provide exciting opportunities for future work to explore the role of the thalamus, and more broadly, the entire limbic memory circuit in AD. As the field of connectomics is rapidly expanding and improving, rapid and reliable segmentation of the thalamus will provide invaluable information necessary for better understanding thalamic connections to cortical and subcortical regions of the brain. We believe that mapping the entire limbic memory circuit using a multimodal approach including diffusion MRI, functional MRI and structural MRI, will provide a much more comprehensive insight into the neural basis of cognitive changes in AD.

In conclusion, this work highlights the importance of considering individual thalamic nuclei in the progression of AD. The thalamus is a complex structure with widespread connections throughout the brain, and one might expect different thalamic nuclei to be affected differently than others as AD progresses. With recent advances in thalamic parcellation algorithms, we can provide higher degrees of sensitivity and specificity to changes within the thalamus, and provide more insight into the progression of AD. While the hippocampus and other parts of the medial temporal lobe are clearly the most strongly affected regions of the brain in AD, the data provided in this work supports the hypothesis that the episodic memory loss that characterize the early stages of the disease may be mediated, at least in part, by the anterior thalamic components of the Papez circuit and the MD nucleus.