Prior‐guided individualized thalamic parcellation based on local diffusion characteristics

Abstract

Comprising numerous subnuclei, the thalamus intricately interconnects the cortex and subcortex, orchestrating various facets of brain functions. Extracting personalized parcellation patterns for these subnuclei is crucial, as different thalamic nuclei play varying roles in cognition and serve as therapeutic targets for neuromodulation. However, accurately delineating the thalamic nuclei boundary at the individual level is challenging due to intersubject variability. In this study, we proposed a prior‐guided parcellation (PG‐par) method to achieve robust individualized thalamic parcellation based on a central‐boundary prior. We first constructed probabilistic atlas of thalamic nuclei using high‐quality diffusion MRI datasets based on the local diffusion characteristics. Subsequently, high‐probability voxels in the probabilistic atlas were utilized as prior guidance to train unique multiple classification models for each subject based on a multilayer perceptron. Finally, we employed the trained model to predict the parcellation labels for thalamic voxels and construct individualized thalamic parcellation. Through a test–retest assessment, the proposed prior‐guided individualized thalamic parcellation exhibited excellent reproducibility and the capacity to detect individual variability. Compared with group atlas registration and individual clustering parcellation, the proposed PG‐par demonstrated superior parcellation performance under different scanning protocols and clinic settings. Furthermore, the prior‐guided individualized parcellation exhibited better correspondence with the histological staining atlas. The proposed prior‐guided individualized thalamic parcellation method contributes to the personalized modeling of brain parcellation.

Based on the “central‐boundary prior” hypothesis and “one person one model” principle, we proposed a prior‐guided individualized thalamic parcellation method to automatically extract the thalamic parcellation pattern, capturing both high individual specificity and population commonality.

1. INTRODUCTION

The dorsal thalamus, commonly referred to as the thalamus, constitutes the largest part of the diencephalon and can be broadly divided into anterior, lateral, and medial nuclei by the internal medullary laminae. These nuclei exhibit not only structural differentiation but also fulfill distinct functions (Jankowski et al., 2013; Watanabe & Funahashi, 2012). They have been identified as selectively vulnerable to specific pathological conditions (Braak & Braak, 1991). Specific thalamic nuclei serve as neuromodulation targets for nervous system diseases (Whiting et al., 2018). For instance, deep brain stimulation of the ventral intermediate nucleus has proven effective in inhibiting essential tremors (Benabid et al., 1991), while the anterior thalamic nucleus is a targeted intervention for intractable epilepsy (Krishna et al., 2016; Kundu et al., 2022). The ventral posterolateral nucleus is a therapeutic target for alleviating neuropathic pain (Abreu et al., 2017), and the centromedian‐parafascicular‐complex is involved with consciousness recovery after severe traumatic brain injury (Schiff et al., 2007) and treatment of Tourette Syndrome (Baldermann et al., 2021). Nevertheless, the therapeutic response varies among patients, suggesting that the thalamic structure of each individual possesses additional, yet unidentified, particular specificity.

The human thalamic nuclei play diverse roles in various brain disorders, and the direct tracing of the boundaries of each thalamic nucleus is typically achieved through histological atlases (Krauth et al., 2010; Morel, 2007; Morel et al., 1997). While histology is often considered the gold standard, discrepancies arising from different anatomists, various histological staining methods, and the limited availability of specimens for creating histological atlases have resulted in a lack of a consistent and unified understanding of the anatomical delineation of the thalamic nuclei (Mai & Majtanik, 2018). MRI‐based thalamus atlases offer a noninvasive, automated, and reliable means to extract the structural or functional characteristics of the thalamus. The parcellation of the thalamus through imaging can be categorized into three types obtained by different MRI modalities: structural MRI (Deoni et al., 2005; Iglesias et al., 2018; Su et al., 2019; Traynor et al., 2011), diffusion MRI (dMRI) (Battistella et al., 2017; Behrens et al., 2003; Fan et al., 2016; Jakab et al., 2012; Jonasson et al., 2007; Kumar et al., 2015; Lambert et al., 2017; Najdenovska et al., 2018; O'Muircheartaigh et al., 2011; Wiegell et al., 2003; Zheng et al., 2021), and functional MRI (Kim et al., 2013; Kumar et al., 2017; Tian et al., 2020; Zhang et al., 2008). Thalamic parcellations based on structural MRI rely on manual labeling priors, whereas parcellations based on dMRI or functional MRI apply data‐driven methods using global connectivity information or local characteristics (Iglehart et al., 2020). While functional connectivity (FC) or structural connectivity (SC) based thalamic parcellation can provide global connectivity information, they identify groups of nuclei corresponding to specific functional circuits (Behrens et al., 2003; Zhang et al., 2008). Early structural‐based thalamic parcellation directly registered the histological atlas (Krauth et al., 2010; Morel, 2007; Morel et al., 1997) to the structural image of the subject without intrinsic contrast. With the introduction of new contrasts (like white matter [WM] null) (Su et al., 2019; Tourdias et al., 2014) and improved Bayesian methods (Iglesias et al., 2018), structural‐based parcellations utilized the intrinsic local structural variation contrast and had been widely used in clinical studies (Schiff et al., 2023; Tian et al., 2023). Complementally, the diffusion‐based parcellations are advantageous in depicting the intrinsic local diffusion orientation contrast of thalamic nuclei (Battistella et al., 2017; Kumar et al., 2015; Wiegell et al., 2003; Zheng et al., 2021).

In addition to the diversity in the parcellation features, the individualized parcellation methods are also upgrading. THOMAS is a multi‐atlas segmentation method for nuclei label assignment based on the similarity between individual features and manual labeling prior (Su et al., 2019). Statistical shape model is another way to integrate the manual contours prior with the individual features (Liu et al., 2020). Since these methods combine individual features and manual prior for thalamic parcellation, they can provide parcellation pattern by integrating individual specificity and population commonality, but with limited prior labeling data due to the high consumption of manual annotation. Generally, current approaches to individualized brain parcellation can be broadly categorized into three classes: group atlas registration (GA‐reg; Evans et al., 2012), unsupervised learning‐based parcellation (Eickhoff et al., 2018), and prior‐guided parcellation (PG‐par; Cui et al., 2020; Kong et al., 2019; Li et al., 2023; Ma et al., 2022; Salehi et al., 2018; Wang et al., 2015). GA‐reg methods assume that different subjects follow the same parcellation pattern, emphasizing high population commonality but overlooking individual specificity. Unsupervised learning‐based parcellation methods leverage individual features to capture individual specificity but are affected by the initialization of the unsupervised learning algorithms and are sensitive to MRI data quality. PG‐par methods aim to construct individualized brain parcellation by integrating individual features in native space with prior guidance from group‐level parcellation patterns, thereby capturing both individual specificity and population commonality. However, prior‐guided individualized parcellation methods can be further refined to achieve more personalized modeling.

In this study, we introduce a prior‐guided individualized thalamic parcellation method. Initially, we constructed probabilistic atlas of thalamic nuclei, identifying high‐probability regions to serve as prior guidance. Subsequently, we trained a classification model using the subject's unique features and applied the trained model to perform personalized thalamic parcellation for the individual. Through test–retest examination, our proposed prior‐guided individualized thalamic parcellation method exhibited exceptional capacities in capturing both intrasubject consistency and intersubject variability. Evaluation across multiple datasets with diverse scanning protocols revealed that our method outperformed GA‐reg and individual clustering in terms of parcellation performance and histological consistency.

2. MATERIALS AND METHODS

2.1. Datasets

In this study, we utilized three publicly available datasets: the Human Connectome Project (HCP) (Van Essen et al., 2013) (https://humanconnectome.org), Multisite, Multiscanner, and Multisubject Acquisitions for Studying Variability (MASiVar) (Cai et al., 2021) (https://openneuro.org/datasets/ds003416/versions/2.0.2), and Allen Brain Reference Atlases (ABRA) (Ding et al., 2016) (https://atlas.brain-map.org).

We extracted five subsets from the HCP S1200 dataset. The HCP unrelated 100 dataset (3 T dMRI, 1.25 mm isotropic, b = 1000, 2000, 3000) was formed using 100 unrelated subjects, and it was employed to construct the probabilistic thalamus atlases. The other four subsets had no overlap with the HCP unrelated 100 and were used to assess the individualized thalamic parcellation method. The HCP test 30 (3 T dMRI, 1.25 mm isotropic, b = 1000, 2000, 3000) and HCP Retest 30 (3 T dMRI, 1.25 mm isotropic, b = 1000, 2000, 3000) consisted of the same 30 subjects with scan‐rescan data and were used to examine the reproducibility of the individualized thalamic parcellation method. HCP 3 T 100 (3 T dMRI, 1.25 mm isotropic, b = 1000, 2000, 3000) and HCP 7 T 100 (7 T dMRI, 1.05 mm isotropic, b = 1000, 2000) were formed using the same 100 subjects and were utilized to assess the robustness of the individualized thalamic parcellation method under different scanning protocols. The HCP 3 T 100 dataset had no overlap with HCP Test 30. The subject lists and information of these five HCP datasets can be found in Supplementary Sheet 1. The scanning protocols for these HCP datasets are available at https://humanconnectome.org/study/hcp‐young‐adult/document/1200‐subject‐data‐release. The HCP S1200 dataset had been preprocessed using the HCP minimal preprocessing pipeline (Glasser et al., 2013).

Additionally, four subjects from the second cohort of the MASiVar dataset (3 T dMRI, 2.5 mm isotropic, b = 1000, 1500, 2000, 2500) were included to assess the robustness of the individualized thalamic parcellation method. Each subject in the MASiVar cohort II dataset underwent four sessions of dMRI acquisition on four different MRI scanners. The scanning protocol for this dataset can be found at https://openneuro.org/datasets/ds003416/versions/2.0.2, and the MRI data had been preprocessed with PreQual v1.0.0 (https://github.com/MASILab/PreQual) under default settings.

The ABRA dataset comprises data from a 34‐year‐old adult with both dMRI data (3 T dMRI, 0.9 mm isotropic, b = 3686) and Nissl staining data. A histological thalamus atlas (Ding et al., 2016) has been manually labeled based on the Nissl staining data. In this study, the MRI data were utilized to construct individualized thalamic parcellation and assess histological consistency. Detailed information regarding the scanning protocol and data preprocessing of the ABRA dataset can be found at www.brainspan.org/static/atlas.

In addition to the publicly available datasets mentioned above, we included data from our laboratory to assess the clinical feasibility of the prior‐guided individualized thalamic parcellation method. This dataset, named NC10, consists of 10 healthy subjects (4 females, mean age: 23.92, standard deviation of age: 4.1) with the single‐shell acquisition of dMRI (3 T dMRI, 2 mm isotropic, b = 1000, 60 volumes). This dataset is not publicly available due to privacy concerns.

2.2. Preprocessing

The data preprocessing followed previous ODF‐based thalamic parcellation methods (Battistella et al., 2017; Najdenovska et al., 2018; Zheng et al., 2021). Morel's histological thalamus atlas (Morel, 2007) was employed as the initial thalamic ROI and registered to native diffusion space using FSL (Jenkinson et al., 2012). As illustrated in Supplementary Figure S1a, the diffusion space (B0) was initially linearly registered into structural space (T1w). Subsequently, the structural space underwent linear and nonlinear registration to MNI space. Finally, the registration warp field from MNI space to diffusion space was generated by converting and inverting these two deformation fields. As depicted in Supplementary Figure S1b, to eliminate the WM and cerebrospinal fluid (CSF), the fractional anisotropy (FA) map was extracted from DWI images using FSL, and the CSF probability map (PM) was calculated from T1w images by SPM (Penny et al., 2011) and registered into diffusion space. Finally, in native diffusion space, voxels in the thalamic ROI with an FA value higher than 0.6 or a CSF probability higher than 0.05 were removed based on the priors (Battistella et al., 2017; Najdenovska et al., 2018; Zheng et al., 2021). An eighth‐order ODF model was employed to characterize the local diffusion model of thalamic nuclei and was estimated in MRtrix3 (Tournier et al., 2019). Consequently, the local diffusion model of each thalamic voxel was represented by 45‐dimensional spherical harmonic (SH) function coefficients. The dhollander (Dhollander et al., 2019) algorithm was used to compute the response function. The single‐shell single‐tissue constrained spherical deconvolution (SSST‐CSD) (Tournier et al., 2008) or multi‐shell multi‐tissue CSD (MSMT‐CSD) (Jeurissen et al., 2014) method was adopted to estimate the eighth‐order ODF model of the thalamic voxels based on the number of b‐values (SSST‐CSD for a single b‐value and MSMT‐CSD for multiple b‐values).

2.3. Individual clustering

The individual clustering scheme was similar to Battistella et al.'s (2017) works. The similarity coefficient between a voxel pair in thalamic ROI was calculated based on the SH coefficients and spatial coordinates according to priors (Battistella et al., 2017; Najdenovska et al., 2018; Zheng et al., 2021). The SH coefficients served to depict the diffusion features, while the spatial coordinates ensured the continuity of the parcellations. The similarity coefficient between a voxel pair was calculated by Equation (1).

$$S\left(i,j\right)=\frac{w_{\mathit{odf}}}{1+\beta E_{\mathit{odf}}\left(i,j\right)}+\frac{1-w_{\mathit{odf}}}{1+E_{\mathit{pos}}\left(i,j\right)}$$ (1)

In this equation, $S\left(i,j\right)$ is the similarity coefficient between voxel $i$ and $j$; $E_{\mathit{odf}}\left(i,j\right)$ is the Euclidean distance of the SH coefficients, and $E_{\mathit{pos}}\left(i,j\right)$ is the Euclidean distance between the spatial coordinates. To ensure an equal contribution of the two features to the final clustering, the weighted coefficient of the SH distance $w_{\mathit{odf}}$ was set to 0.5 (Battistella et al., 2017). To avoid bias in the clustering (Battistella et al., 2017; Najdenovska et al., 2018; Zheng et al., 2021), the SH coefficients were scaled by multiplying them by a scaling factor $\beta$ to ensure that these two distances were in the same scale. The scaling factor $\beta$ was empirically set to 89, 89, 98, 30, 108, and 55 when processing the HCP Unrelated 100, HCP 3T 100, HCP 7T 100, MASiVar, ABRA, and NC10 datasets. Finally, spectral clustering was applied in each hemisphere to cluster the thalamic voxels with a similar local diffusion pattern. The potential cluster number K was set in the range of 2–28 to ensure an intrinsic parcellation pattern.

2.4. Probabilistic atlas construction

All individual clustering results were registered from native diffusion space to MNI space to construct the probabilistic thalamic atlas like Najdenovska et al.'s (2018) work. Initially, a group reference map was built to establish the labeling criterion. An N‐dimensional (N = 100) label distribution vector, representing the voxel‐wise labeling pattern among these 100 participants, was computed for each thalamic voxel (Zhang et al., 2015). The group reference map was generated by computing the correlation on the label distribution vectors of all the thalamic voxels and performing spectral clustering on the correlation matrix. Subsequently, employing the maximum spatial overlapping (Munkres, 1957), cluster labels between the participants and hemispheres were uniformly and symmetrically relabeled. Next, the voxel‐wise PM of the thalamus was constructed by calculating the probability of any given voxel belonging to the corresponding K clusters among all relabeled individual clustering results. Finally, by applying a probability threshold (25%) on the PM and assigning cluster labels with the maximum probability to the surviving voxels, the maximum PM (MPM) of the thalamus could be constructed. The entire pipeline for constructing the probabilistic atlas is illustrated in Figure 1a. Visualization of the parcellation results was performed using ITK‐SNAP (Yushkevich et al., 2006). The nomenclature of thalamic subregions followed previous ODF‐based thalamic parcellation methods (Battistella et al., 2017; Najdenovska et al., 2018).

FIGURE 1.

Pipeline of the prior‐guided individualized thalamic parcellation method. (a) Pipeline of probabilistic atlas construction. Local diffusion characteristics are extracted in the individual space to calculate the similarity matrix for clustering. Then, the clustering results of 100 unrelated subjects are registered to the MNI space to construct the probabilistic atlas of the thalamic nuclei. (b) Pipeline of prior‐guidance extraction. Areas with high probability values in the probabilistic atlas are extracted and registered to the individual space as the individual prior parcellation to provide prior guidance for individualized parcellation. (c) Pipeline of prior‐guided individualized thalamic parcellation. Individual prior parcellation is used to train a classification model to predict the cluster labels for unlabeled thalamic voxels. Merging the prior parcellation with the unlabeled parcellation to construct the individualized parcellation.

2.5. Optimal parcellation granularity estimation

Two robust and conclusive indexes were computed to determine the intrinsic optimal cluster numbers of the thalamic parcellation like Fan et al.'s (2016) work. The Dice coefficient (Dice, 1945) (Dice) was employed to measure the intersubject consistency of thalamic parcellation. A leave‐one‐out strategy was applied to divide the HCP unrelated 100 dataset into 1‐subject and 99‐subject datasets. The mean Dice coefficient was calculated over 100 repetitions of the leave‐one‐out strategy, considering cluster numbers from 2 to 28. In each repetition, the Dice coefficient was computed between a temporary MPM constructed by 99‐subject dataset and the individual clustering result by the 1‐subject dataset. The Dice coefficient equation is presented in Equation (2).

$$\mathit{\text{Dice}}\left(K\right)=\frac{1}{K}\sum _{k=1}^K\frac{2*\mid A_k\cap B_k\mid }{\mid A_k\mid \cup \mid B_k\mid }$$ (2)

In this equation, $K$ represents the number of clusters; $A_k$ or $B_k$ represents the kth of $K$ clusters in parcellation A or B; $\mid A_k\cap B_k\mid$ represents the number of spatially overlapping voxels between $A_k$ and $B_k$; $\mid A_k\mid \cup \mid B_k\mid$ represents the total number of voxels in $A_k$ and $B_k$. A higher Dice coefficient signifies greater intersubject consistency, indicating a potential population‐level parcellation pattern of the thalamus.

Topological distance (Tungaraza et al., 2015) (TpD) was adopted to quantify the homology between two hemispheres of the MPM across cluster numbers from 2 to 28. The $\mathit{TpD}$ equation is presented in Equation (3).

$$\mathit{TpD}\left(K\right)=\mathit{cos}\left(M_K^l,M_K^r\right)$$ (3)

In this equation, $M_K^l$ and $M_K^r$ are both two‐dimensional matrices representing the left and right intrahemispheric neighboring patterns under $K$ clusters, respectively. $M_K^l\left(i,j\right)$ and $M_K^r\left(i,j\right)$ represent the number of neighboring voxels of cluster $i$ and cluster $j$ in the left hemisphere and right hemisphere, respectively. The cosine distance was calculated after vectorizing these two matrices to obtain the TpD coefficients. A lower topological distance represents higher interhemispheric consistency. Consistent with prior literatures (Fan et al., 2016; Li et al., 2022), the intrinsic optimal cluster numbers were identified at the local extremes of these two indexes.

2.6. Prior‐guidance extraction

In the probabilistic atlas, voxels with high probability values represent parcellation patterns that are common among the subjects, while voxels with low probability values represent parcellation variabilities between subjects. When increasing the probability threshold in the probabilistic thalamus atlas, boundary voxels of the thalamic nuclei are gradually eliminated, leaving the central area intact. Based on this observation, we formulate the “central‐boundary prior” hypothesis. The boundaries of subregions in the probabilistic atlas primarily reflect parcellation variabilities between subjects, whereas the central area of subregions in the probabilistic atlas could mainly reflect the parcellation patterns common among subjects. Based on this “central‐boundary prior,” we nonlinearly registered the central area of subregions (high probability, $p\ge 80\%$) in probabilistic atlas from MNI space to native diffusion space. Then, we used the resulting area as an individual prior parcellation (seen in Figure 1b). This individual prior parcellation was then utilized for training the classification model for prior‐guided individualized parcellation.

2.7. Prior‐guided individualized thalamic parcellation

We train an individualized parcellation model for each subject based on a multilayer perceptron network. The network consists of a 48‐dimensional input layer, two m‐dimensional intermediate layers (m = 48 in this work), and a k‐dimensional output layer (k = 7 in this work). The activation function between the first three layers is a rectified linear unit, while we used a softmax activation function between the third and output layers to ensure that outputs can be interpreted as probabilities. The training set for each subject is the individual prior parcellation extracted from the probabilistic atlas, while the test set is the voxels in thalamic ROI but outside the individual prior parcellation (unlabeled voxels). The input feature is a 48‐dimensional vector with 45 ODF coefficients and 3 spatial coordinates. The output is given as a k‐dimensional (k = 7 in this work) one‐out‐of‐k encoded vector corresponding to the probabilities of belonging to each of the k clusters. The objective function is based on the cross‐entropy loss between the nuclei label in individual prior parcellation and model predicted label. For optimization, we use adaptive moment estimation with an initial learning rate of 0.005. The network is regularized by random dropout with a rate of 0.4. Additional parameters included a batch size of 32 and a maximum number of epochs of 100. Using the trained classifier, the K clusters' probability vector of the unlabeled voxels can be predicted, and the cluster label with the maximum probability is assigned to the individual unlabeled parcellation. The final individualized parcellation is constructed by merging the prior parcellation and the unlabeled parcellation. Individualized thalamic parcellations of each hemisphere and each subject are constructed independently. Figure 2c illustrates the pipeline of prior‐guided individualized thalamic parcellation.

FIGURE 2.

Probabilistic atlas of the thalamus based on local diffusion characteristics. (a) Local maxima in the Dice coefficient were observed in both the left and right hemispheres when the number of clusters was set to 7, and the Tpd coefficient approached zero. (b) The voxel count for each thalamic nucleus progressively decreased with increasing probability threshold. (c) With the rise in probability threshold, the boundary voxels of thalamic nuclei were gradually eliminated, while the central area was retained.

2.8. Test–retest examination

HCP Test 30 and HCP Retest 30 were employed to examine the reproducibility of the prior‐guided individualized thalamic parcellation. Initially, we registered the prior‐guided thalamic parcellations of all subjects into the MNI space. Subsequently, we calculated the dice coefficient (Dice, 1945) for each subject pair using the equation in Equation (2). Intrasubject consistency was computed by averaging the dice coefficient between the test and retest thalamic parcellations for the same subject (30 subjects in total). Meanwhile, in the MNI space, intersubject consistency was computed by the thalamic parcellations between subject pairs in each dataset (a total of 870 subject pairs on two datasets). Intrasubject consistency was employed to illustrate the test–retest consistency, while intersubject consistency was used to quantify the variability between thalamic parcellations across subjects.

2.9. Robustness assessment

The HCP 3T 100, HCP 7T 100, MASiVar, and NC10 datasets were employed to assess the robustness of the prior‐guided individualized thalamic parcellation under different scanning protocols. Cohesion within and separation between clusters are the criteria for evaluating clustering goodness (Chen et al., 2012). The clustering goodness was quantified using the silhouette coefficient (Sil) in Equation (4).

$${\mathit{Sil}}_i=\frac{\left(b_i-a_i\right)}{\mathit{max}\left(a_i,b_i\right)}$$ (4)

In this equation, $a_i$ represents the mean distance between voxel $i$ and voxels in the same cluster, while $b_i$ represents the mean distance between the voxel $i$ and voxels in other clusters. The distance measurement was based on the Euclidean distance of the 45‐dimensional SH coefficients, which directly depicted the local diffusion features of the thalamic nuclei. A higher silhouette coefficient indicates better parcellation performance. To assess the merits of the prior‐guided parcellation method, a comparison was made with the individual clustering parcellation (IC‐par) method (Battistella et al., 2017) and GA‐reg method (Najdenovska et al., 2018). The IC‐par method constructed the thalamic parcellation by performing clustering in native diffusion space using the same parcellation pipeline described in the individual clustering subsection. The GA‐reg method nonlinearly registered the MPM to native diffusion space and regarded it as the thalamic parcellation for a new subject. In addition to the clustering goodness, we ran THOMAS (Su et al., 2019; Vidal et al., 2024) on T1 data (https://github.com/thalamicseg/hipsthomasdocker) as the silver standard to compare PG‐par with IC‐par and GA‐reg. THOMAS is an automatic multi‐atlas segmentation method for thalamic parcellation and has already been validated against manual segmentation. We registered the thalamic parcellation results of GA‐reg, IC‐par, PG‐par in diffusion space (B0 image) to THOMAS segmentation results in structural space (T1w image) using affine registration by fnirt function in FSL (Jenkinson et al., 2012). On each subject in these four datasets, we then calculated Dice coefficients described in Equation (2) between the THOMAS segmentation results with GA‐reg, IC‐par, and PG‐par, respectively. Finally, on each dataset, we visualized the results of the quantitative comparisons between the three methods for using THOMAS as the silver standard.

2.10. Histological comparison

The ABRA dataset was used to perform histological comparison (Ding et al., 2016). The histological atlas was only published in 2D format on a website https://atlas.brain-map.org. Hence, the histological consistency was examined by visually comparing the automatically constructed thalamic parcellation based on MRI data with the manually labeled reference atlas (Ding et al., 2016) based on Nissl staining data. To demonstrate the histological correspondence, thalamic parcellations were constructed based on the ABRA MRI data using the IC‐par, GA‐reg and PG‐par methods.

3. RESULTS

3.1. Probabilistic atlas of thalamus

We constructed PM and MPM of the thalamus based on the HCP unrelated 100 dataset. To quantify the parcellation consistency across subjects and the homology between hemispheres, we calculated the Dice coefficient and topological distance coefficient for cluster numbers ranging from 2 to 28. As depicted in Figure 2a, when the cluster number was set to 7, the Dice coefficient exhibited local maxima in both the left and right hemispheres, while the TpD coefficient approached to zero. This suggests that 7 is potential optimal cluster numbers for capturing the local diffusion‐based parcellation pattern of thalamic nuclei. Detailed results of Dice and TpD can be seen in Supplementary Sheet 2. Figure 2b demonstrates that the number of voxels in all the thalamic nuclei uniformly decreased with increasing probability threshold. At a probability threshold of 25%, the thalamus MPM exhibited a homologous parcellation pattern between the two hemispheres and a high level of spatial continuity within each hemisphere (seen in Figure 2c). Moreover, with the rising probability threshold, the boundary voxels of the thalamic nuclei were gradually eliminated, while the central areas were retained. Based on this observation, we formulated the “central‐boundary prior” hypothesis. That is, the boundary of subregions in probabilistic atlas primarily reflects parcellation variabilities between subjects, whereas the central area of subregions in probabilistic atlas predominantly reflects the parcellation pattern common among subjects.

3.2. Test–retest examination results

The HCP Test 30 and HCP Retest 30 datasets were employed to examine the reproducibility of the prior‐guided individualized thalamic parcellation method. Dice coefficients for the prior‐guided thalamic parcellations in MNI space were computed to quantify intrasubject consistency and intersubject variability. As depicted in Figure 3a, the intrasubject Dice was significantly higher (p < .001) than the intersubject Dice. Detailed intrasubject and intersubject Dice coefficients for the test–retest examination can be found in Supplementary Sheet 3. These results underscore the excellent reproducibility of PG‐par and its ability to effectively capture intersubject variability.

FIGURE 3.

Results of the test–retest examination for the prior‐guided individualized thalamic parcellation method. Consistency between test and retest for the prior‐guided individualized thalamic parcellations. The intrasubject consistency is significantly higher than the intersubject one.

3.3. Robustness assessment results

Prior‐guided parcellation was compared to IC‐par (Battistella et al., 2017) and GA‐reg (Najdenovska et al., 2018) methods across multiple datasets. The silhouette coefficient was used to quantify clustering goodness, reflecting the cohesiveness within and separation between clusters. Robustness assessment was conducted on HCP 3T 100, HCP 7T 100, MASiVar, and NC10 datasets with different scanning protocols. Shown in Figure 4, PG‐par demonstrated significantly higher silhouette coefficients than IC‐par and GA‐reg (p < .05). Supplementary Sheet 4 contains the silhouette coefficients for these thalamic parcellation methods. Additionally, shown in Figure 5, compared to IC‐par and GA‐reg, PG‐par provided significantly higher consistency to the silver standard THOMAS method (p < .05). Supplementary Sheet 5 contains the dice coefficients with THOMAS of these thalamic parcellation methods.

FIGURE 4.

Clustering goodness of the individualized thalamic parcellation methods. (a–d) Silhouette coefficients of the GA‐reg, IC‐par, and PG‐par thalamic parcellations on the Human Connectome Project (HCP) 7T 100, HCP 3T 100, MASiVar, and NC10 datasets. The silhouette coefficients of the PG‐par thalamic parcellations are significantly higher than those of the GA‐reg and IC‐par on each dataset. GA‐reg, group atlas registration; IC‐par, individual clustering parcellation; PG‐par, prior‐guided parcellation.

FIGURE 5.

Comparison of accuracy of methods compared to THOMAS structural segmentation. (a–d) Dice with THOMAS of the GA‐reg, IC‐par, and PG‐par thalamic parcellations on the Human Connectome Project (HCP) 7T 100, HCP 3T 100, MASiVar, and NC10 datasets. PG‐par provides significantly higher consistency with THOMAS than GA‐reg and IC‐par on each dataset. GA‐reg, group atlas registration; IC‐par, individual clustering parcellation; PG‐par, prior‐guided parcellation.

3.4. Histological comparison results

Based on the MRI data provided by the ABRA dataset, we generated thalamic parcellations using PG‐par, GA‐reg, and IC‐par, respectively. To assess histological consistency, we qualitatively compared these MRI‐based thalamic parcellations with a manually labeled histological atlas. This histological atlas, constructed through Nissl staining data, serves as the gold standard for depicting thalamic nuclei parcellation patterns. As illustrated in Figure 6a, the PG‐par thalamic parcellation demonstrated a parcellation pattern that was more likely to align with the manually labeled histological atlas compared to those generated by GA‐reg and IC‐par (pointed by the orange arrow). Additionally, as shown in Figure 6a,b, all three methods yielded thalamic parcellation patterns that were similar to the histological atlas.

FIGURE 6.

Histological comparison of the individualized thalamic parcellation methods. (a) The prior‐guided parcellation (PG‐par) exhibits a parcellation pattern that is more likely to align with the histological atlas (pointed by the orange arrow). (b) All three methods yield similar parcellation patterns to the histological atlas.

4. DISCUSSION

In this study, we developed a prior‐guided individualized thalamic parcellation method based on the central‐boundary prior. We examined the reproducibility of this method using a test–retest dataset and found that the intrasubject consistency was significantly higher than the intersubject consistency. We assessed the robustness of this method on multiple scanning protocols and compared it to individual clustering parcellation and group atlas registration. The prior‐guided parcellation demonstrated significantly higher clustering goodness and silver standard consistency compared to group atlas registration and individual clustering. Furthermore, it demonstrated enhanced histological consistency with a manually labeled atlas based on histological data.

Using HCP unrelated 100 dataset (Van Essen et al., 2013), we constructed a probabilistic thalamus atlas following Najdenovska et al.'s (2018) work but with a larger number of subjects and higher‐order ODF characteristics. Higher‐order ODF models rely on more SH functions for estimation, thereby providing a more accurate modeling of the local diffusion pattern. In comparison to previous DTI studies (Jonasson et al., 2007; Kumar et al., 2015; Wiegell et al., 2003) or lower‐order ODF studies (Battistella et al., 2017; Najdenovska et al., 2018; Zheng et al., 2021), the eighth order ODF model can capture a more detailed local diffusion pattern. While more detailed estimation may be affected by the signal‐to‐noise ratio (SNR) of dMRI data, the robustness of eighth order ODF has been evaluated, and the estimation of fiber orientation does not noticeably deviate as the SNR decreases (Jeurissen et al., 2014). ODFs at different orders have been employed in previous literature (Battistella et al., 2017; Najdenovska et al., 2018) and also in this work. Their abilities were similar in capturing diffusion similarity of thalamic nuclei, as observed in Supplementary Figure S2, possibly because ODFs at different orders essentially provide the same estimation of the principal components of the local diffusion pattern, with more SH functions offering additional detail. Furthermore, based on an unbiased ODF template established by averaging ODF data from 100 subjects, we discovered that the eighth order ODF exhibited a stronger ability to quantify intra‐subregion similarity and inter‐subregion variability in thalamic parcellation compared to lower‐order ODFs, as illustrated in Supplementary Figure S3.

To achieve individualized thalamic parcellation, we utilized the intrinsic local diffusion orientation contrast of the thalamic nuclei and group atlas prior. With advances in imaging technology and machine learning algorithms, there are many remarkable methods that can utilize the intrinsic contrast and prior guidance for thalamic parcellation. For example, THOMAS (Su et al., 2019) utilized white‐matter‐nulled sequences to extract the intrinsic structural contrast and manual labeling prior for thalamic parcellation, and had been well used in neuromodulatory localization after severe traumatic brain injury (Schiff et al., 2023). As another example, the thalamic parcellation method in Freesurfer (Iglesias et al., 2018) combined the individual structural contrast and histological atlas prior by Bayesian modeling, and had also been used in clinical studies (Tian et al., 2023). Notably, local diffusion orientation contrast is a complement to local structural variation contrast. They reveal intrinsic parcellation patterns of thalamic nuclei from different perspectives.

The PG‐par demonstrated superior parcellation performance compared to GA‐reg and IC‐par across multiple datasets. While group atlas registration provides high population commonality, it lacks individual specificity. On the other hand, individual clustering can leverage individual features but is sensitive to algorithm initialization and perturbations caused by noise in the data. The prior guided individualized parcellation strategy integrates and adapts group‐level prior information into each individual, striking a balance between high population commonality and individual specificity. Some prior‐guided individualized parcellation methods constructed a probabilistic atlas (Wang et al., 2015), or prior exemplars (Salehi et al., 2018), or group level components (Cui et al., 2020) to initialize the unsupervised learning algorithms, and then applied clustering (Salehi et al., 2018; Wang et al., 2015) or decomposition (Cui et al., 2020) for subsequent individualized parcellation. In this work, similar to Wang et al.'s (2015) method, we constructed a probabilistic atlas for initialization but applied a neural network model for subsequent individualized parcellation. Other attempts to train models for subsequent individualized parcellation utilized a group of subjects, including a multi‐session hierarchical Bayesian model (Kong et al., 2019), graph convolutional network (Ma et al., 2022), and autoencoder model (Li et al., 2023). While these parcellation models (Kong et al., 2019; Li et al., 2023; Ma et al., 2022) are stable across subjects, they still rely on group information as they are trained within a group of subjects. These group‐level models can hardly fully reflect the individual‐level parcellation models. Based on “central‐boundary prior,” PG‐par trains a multilayer perceptron classifier for each subject, utilizing only individual features. This method allows the model to be independent of the contrast in individual subjects and more accurately represent the parcellation pattern of the subject.

We proposed a “central‐boundary prior” hypothesis based on the probabilistic thalamic atlas. According to this hypothesis, central regions in the probabilistic atlas are preserved across the population and represent the population commonality of the parcellation pattern, while the boundary regions reflect individual variability. This “central‐boundary prior” is also supported by previous probabilistic brain atlases based on cytoarchitecture (Amunts et al., 2020), local diffusion characteristics (Najdenovska et al., 2018), WM tracts (Basile et al., 2022), SC (Fan et al., 2016), FC (Gordon et al., 2016), and manually labeling (Pauli et al., 2018). These probabilistic brain atlases suggest that central regions effectively capture parcellation patterns with high population commonality across data modalities and regions of interest. Aligning the central region of high probability values in the probabilistic atlas to individuals not only preserves parcellation patterns with high population commonality but also enables a more purely individualized parcellation modeling. In this way, the “central‐boundary prior” aligns with the one‐person‐one‐model principle for individualized brain parcellation.

The one‐person‐one‐model principle forms the central concept and core idea behind prior‐guided individualized thalamic parcellation. Essentially, this principle advocates that brain parcellation should rely exclusively on individual‐specific parcellation features and pipeline parameters. While group‐optimal parcellation pipeline parameters are stable across subjects, they do not fully represent the individual‐optimal parcellation model. The utilization of individual‐optimal parameters allows for a more precise capture and characterization of individual specificity. In this context, the prior‐guided individualized parcellation method holds promise for applications in disease diagnosis (Dickie et al., 2018; Wang et al., 2020), neurosurgery localization (Lynch et al., 2022; Plantinga et al., 2018), and prognosis prediction (Charlebois et al., 2022; Wang et al., 2021). Given that disease (Yu et al., 2021), development (Bethlehem et al., 2022), and aging (Damoiseaux, 2017) influence the parcellation pattern at both individual and group levels, guidance from population‐specific probabilistic atlases would make more sense for PG‐par.

It is possible to extend the prior‐guided individualized thalamic parcellation approach to other brain regions and different parcellation features. For instance, this approach can be applied to commonly targeted regions for deep brain stimulation in clinical settings, such as the globus pallidus internus and the subthalamic nuclei (Ewert et al., 2018). These regions exhibit distinct local characteristics, fiber projections, and varying volume sizes, posing challenges for feature selection and model optimization. In such cases, global connectivity may be more suitable for these ROIs as they are located within specific functional circuits (Patriat et al., 2018; Plantinga et al., 2018). Additionally, for the parcellation of small nuclei where labeled data are limited, lightweight machine learning algorithms like support vector machine (Noble, 2006) or random forest (Biau & Scornet, 2016) may provide better classification performance compared to more complex algorithms. This is particularly relevant as these algorithms can provide effective solutions when data availability is scarce for smaller nuclei.

5. LIMITATIONS

Despite conducting test–retest examinations and robustness assessments of the PG‐par method across multiple datasets, this study still lacks quantitative histological validation. The ABRA histological thalamus atlas utilized in this study was only published on the website in a 2D format, limiting our comparisons to qualitative histological assessments. Another challenge is the insufficient amount of available histological data (N = 1), which undermines the credibility of histological comparison. While the PG‐par method is in higher agreement with the silver standard THOMAS than the individual clustering and group atlas registration methods, the local diffusion‐based PG‐par is not in very high agreement with the structural‐based THOMAS, partly due to PG‐par lacking histological and manual labeling prior. On the other hand, diffusion‐based thalamic parcellations lack anisotropy in most gray matter thalamic regions and suffer from distortion, leading to discordance with anatomical nuclei. Using the HCP unrelated 100 datasets, we automatically constructed temporary probabilistic thalamus atlases to identify parcellation patterns in healthy adults. Limited by the spatial resolution of echo planar imaging and poor anisotropy of gray matter, the constructed probabilistic thalamus atlas based on local diffusion characteristics is at a coarser parcellation granularitie than histological atlas (Krauth et al., 2010; Morel, 2007; Morel et al., 1997) and manual labeling guided parcellations (Iglesias et al., 2018; Su et al., 2019). Also, the atlases based on healthy adult data may not fully represent parcellation patterns across different ages and pathological states. While we examined the feasibility of PG‐par on clinical‐level data, it was not tested on actual patients. In the absence of a patient‐specific probabilistic atlas, the PG‐par method requires the creation of a new one to provide prior guidance, posing a challenge for clinical applicability. Therefore, creating probabilistic atlases for specific diseases and age groups is imperative and will be the focus of our future endeavors to enhance clinical applicability.

6. CONCLUSION

In summary, building on the concept of the “central‐boundary prior,” we have introduced a prior‐guided individualized thalamic parcellation method. This method exhibits a robust capability to extract the thalamic parcellation pattern, capturing both high individual specificity and population commonality. The adoption of this prior‐guided individualization strategy holds the potential to facilitate and inspire the brain mapping community.

CONFLICT OF INTEREST STATEMENT

The authors declare that the research was conducted without any commercial or financial relationships that could be construed as a potential conflict of interest.

Supporting information

FIGURE S1. ROI registration and postprocessing. (a) ROI registration. First, the B0 image (native diffusion space) was linearly registered into the T1w image (native structural space) by the flirt function in FSL. Next, the T1w image was linearly and nonlinearly registered into MNI space by the flirt and fnirt function in FSL. Then, by converting the warp matrix of B0_2_T1_WARP and T1_2_MNI_WARP, the warp matrix registering native diffusion space to MNI space could be generated. Finally, by inverting the warp matrix of B0_2_MNI_WARP, the warp matrix registering MNI space into native diffusion space could be generated. (b) ROI postprocessing. First, we extracted the FA map from the DWI images by the dtifit function in SPM. Next, we extracted the CSF probability map from the T1w image by the segmentation function in SPM and registered it to the native diffusion space by flirt in FSL. Then, we registered the thalamic ROI of Morel's thalamus atlas from MNI space to native diffusion space. Finally, we removed the thalamic voxels with an FA value higher than 0.6 and a CSF probability higher than 0.05 to generate the ROI in native diffusion space.

FIGURE S2. Order similarity between ODFs with orders from 4th to 12th. (a) Mean order similarity matrix. (b) Mean order similarity along with the increase of ODF order. (c) The standard deviation of order similarity along with the increase of ODF order. (d) Computation time of ODFs with different orders.

We conducted a comparison of the ability to characterize intrinsic diffusion similarity among thalamic voxels at higher orders. The quantitative results indicate that the 8th order ODF is more suitable for depicting intrinsic diffusion similarity compared to orders 4th, 6th, 10th, and 12th. Initially, we randomly sampled 10 subjects from the HCP unrelated 100 dataset. We computed ODF similarity matrices (voxel by voxel) for thalamic ROI from the 4th to the 12th order for each subject. Subsequently, we computed the order similarity matrix (order by order) between these ODF similarity matrices. Finally, we averaged the order similarity matrices for all sampled subjects, and the result is depicted in Figure S3a. Mean order similarity was calculated by averaging each column in Figure S4a, revealing that the ability of ODFs at different orders to capture intrinsic diffusion similarity is very close (mean order similarity of all orders higher than 0.92 in Figure S4b). It is also noteworthy that the 8th order exhibits higher mean order similarity compared to the other five orders. Additionally, we calculated the standard deviation of each column in Figure S4a, and the 8th order demonstrated a lower standard deviation (as shown in Figure S4c). Furthermore, we compared the computation time of ODFs with different orders for one HCP subject using a single core of the Intel GOLD 6330 CPU. The computation time increases significantly when the order of ODFs exceeds the 8th (as seen in Figure S4d).

FIGURE S3. Intra‐region similarity divided by inter‐region similarity ratio in different cluster numbers with ODF order from 2nd to 8th. The 8th order ODF consistently demonstrates a higher intra‐region similarity/inter‐region similarity ratio with an increasing number of clusters. This indicates that the 8th order ODF exhibits better parcellation performance compared to the other three orders in this analysis.

We conducted a quantitative evaluation of the parcellation similarity, employing the intra−/inter‐region similarity ratio, for four different orders ranging from 2nd to 8th. The 8th order ODF demonstrated a higher intra‐region similarity to inter‐region similarity ratio with an increase in the cluster number, suggesting superior parcellation performance within these four orders.

The thalamic parcellation utilized in this analysis was constructed using the HCP unrelated 100 datasets, the same dataset employed in our manuscript. Using this identical unrelated 100 dataset, ODFs with different orders (2nd, 4th, 6th, 8th) were computed in each subject's diffusion space, registered into MNI space, and averaged to create ODF templates. Consequently, we established 100‐subject thalamic parcellation templates ranging from 2 to 28 clusters and 100‐subject ODF templates for orders 2nd, 4th, 6th, and 8th. Subsequently, using each ODF template, we computed the intra−/inter‐region similarity ratio for each parcellation template from 2 to 28 clusters.

The intra‐region similarity is defined as the mean Pearson correlation of ODF coefficients between each voxel pair within the same subregion, while the inter‐region similarity is the mean Pearson correlation of ODF coefficients between each voxel pair from different subregions. Higher intra‐region similarity and lower inter‐region similarity indicate better parcellation performance. We then calculated the ratio of intra‐region similarity to inter‐region similarity to directly indicate parcellation performance, where a higher intra‐region similarity to inter‐region similarity ratio signifies better parcellation performance. As depicted in Figure S4, the 8th order ODF consistently exhibited a higher intra‐region similarity/inter‐region similarity ratio across varying numbers of clusters, indicating that the 8th order ODF outperforms the other three orders in terms of parcellation performance.

DATA S1: Supporting Information.

DATA S2: Supporting Information.

DATA S3: Supporting Information.

DATA S4: Supporting Information.

DATA S5: Supporting Information.

Gao, C. , Wu, X. , Wang, Y. , Li, G. , Ma, L. , Wang, C. , Xie, S. , Chu, C. , Madsen, K. H. , Hou, Z. , & Fan, L. (2024). Prior‐guided individualized thalamic parcellation based on local diffusion characteristics. Human Brain Mapping, 45(4), e26646. 10.1002/hbm.26646

DATA AVAILABILITY STATEMENT

The HCP S1200 dataset is available at https://humanconnectome.org. The ABRA dataset is available at https://atlas.brain-map.org. The MASiVar dataset is available at https://openneuro.org/datasets/ds003416/versions/2.0.2. The FSL can be downloaded at https://fsl.fmrib.ox.ac.uk/fsl/. The SPM can be downloaded at https://www.fil.ion.ucl.ac.uk/spm/. The MRtrix3 can be downloaded at https://www.mrtrix.org. The ITK‐SNAP can be downloaded at http://www.itksnap.org. The HIPS‐THOMAS is available at https://github.com/thalamicseg/hipsthomasdocker.