Automated Surface‐Based Segmentation of Deep Gray Matter Regions Based on Diffusion Tensor Images Reveals Unique Age Trajectories Over the Healthy Lifespan

ABSTRACT

Many studies have demonstrated unique trajectories of deep gray matter (GM) volumes over development and aging, suggesting but not measuring microstructural alterations. Only a few studies have measured diffusion tensor imaging (DTI) parameters in deep GM or reported these values across a wide age range in a large cohort. Because many clinical protocols do not acquire T1‐weighted images and registration between T1‐weighted and diffusion images is error‐prone, an automated segmentation technique is proposed that works solely on parametric maps calculated from DTI, enabling efficient DTI studies of deep GM in large cohorts without the need for additional structural imaging. The algorithm segments five subcortical GM structures by deforming 3D models to their boundaries visible on diffusion‐weighted images and DTI maps. This DTI‐only method is compared against standard T1‐weighted segmentation using a 1.5‐mm isotropic test–retest diffusion data (n = 24). Inter‐method Dice coefficients were high (> 0.7) for 5 of 10 structures but were low for the left/right globus pallidus, left/right amygdala, and right hippocampus. The proposed DTI‐only segmentation qualitatively appeared more accurate and yielded smaller volumes than T1w for all 10 structures. The segmentation method was then applied to a “lifespan” cohort (n = 357, 5–90 years, 203 females) to assess age changes in volume, fractional anisotropy (FA), and mean diffusivity (MD). For all five structures, MD trajectories were quadratic, decreasing and then increasing after ~30–35 years. For the globus pallidus and hippocampus, FA trajectories remained flat from 5 to ~25 years and then started to decrease over 5–90 years; FA decreased linearly for amygdala, increased linearly for striatum, and remained constant for the thalamus. Notably, the trajectories for DTI were distinct from those of the deep GM volumes. The proposed automated deep GM segmentation method on DTI‐only will facilitate the analysis of deep GM DTI (not typically measured in most studies) and will be advantageous for studies without a T1‐weighted scan, as in many clinical populations.

This paper introduces a novel, automated deep gray matter segmentation method that relies exclusively on diffusion‐weighted imaging, eliminating the need for T1‐weighted images. When applied to a lifespan cohort from ages 5 to 90, the technique successfully mapped distinct, nonlinear developmental and aging trajectories for microstructural diffusion parameters. Ultimately, this efficient approach greatly facilitates DTI analysis in large‐scale studies lacking structural scans.

1. Introduction

Magnetic resonance imaging (MRI) studies of subcortical/deep gray matter (GM) structures (e.g., globus pallidus, caudate nucleus, nucleus accumbens, putamen, thalamus, hippocampus, and amygdala) have primarily focused on their volume (measured on T1 scans) that show unique nonlinear cross‐sectional trajectories over the healthy lifespan that increase until adolescence, peaking between ~15 and 25 years of age, and thereafter decrease with aging (e.g., [1, 2, 3]. There are far fewer investigations of deep GM using diffusion tensor imaging (DTI) that have provided complementary parameters indicative of region‐dependent microstructural changes of multiple deep GM regions with typical child development to young adulthood (e.g., [4, 5, 6, 7, 8]) and aging (e.g., [9, 10, 11, 12]). In contrast to DTI of white matter (WM), a primary reason for so few DTI studies of deep GM in large healthy populations is the limitations of commonly used analysis methodologies, such as manual region‐of‐interest (ROI) segmentation that is time consuming. For practical limitations, such manual ROI studies then usually extract measurements from only a portion of the structures such as on a single slice.

There are two main automated approaches to segment the deep GM to extract diffusion parameters. The first approach uses a set of pre‐segmented ROIs in a standard diffusion template space (often age‐relevant) that is then applied to parcellate brain regions in coregistered diffusion images and maps within individual participants [13, 14]. However, this relies on proper registration that can be problematic, particularly for small deep GM regions, given differences in the anatomical shape and image acquisition (e.g., in‐plane resolution, slice thickness, and image distortions). Furthermore, the DTI templates (e.g., ICBM‐DTI‐81) are usually focused on parcellations of WM regions [15]. The second approach involves identifying voxels belonging to deep GM regions using one of many automated methods on an additionally acquired T1‐weighted (T1w) structural image that is then coregistered to the diffusion images [7]. However, this registration is prone to errors because the diffusion images acquired with single‐shot EPI typically have spatial distortions not present in the structural image and are usually acquired at a much lower spatial resolution. In addition, basing segmentation techniques on T1 weighting alone is problematic because several boundaries of deep GM structures have poor WM/GM contrast on T1w images, notably the lateral border of the thalamus adjacent to the internal capsule, the borders of the globus pallidus adjacent to the internal and external capsules, and the lateral border of the putamen adjacent to the external capsule. This lack of contrast limits the use of deformable mesh‐based segmentation methods that use the 3D geometry of the target structure to align the edges of a 3D model to regions with large intensity gradients on an image, as has been developed for cortex [16, 17, 18] and subcortical structures [19]. To employ deformation‐based segmentation methods, FA maps, which have excellent WM/GM contrast in regions surrounding the basal ganglia, have been proposed to improve mesh‐based segmentation compared to methods using a single image contrast [20]. Given that the other boundaries of the subcortical GM are also visible on diffusion‐weighted images (DWI) and maps, this suggests that diffusion MRI alone may provide the contrast necessary to segment the deep GM structures as has been achieved recently for the cortex [21], avoiding the need for additional structural imaging, or problematic registration between different image types.

To date, only four studies have developed deep GM segmentation methods that work on native diffusion images with techniques relying on deep learning. The first approach used diffusion parameter maps (axial diffusivity, mean diffusivity, radial diffusivity, and apparent fiber density) as input to learn a 10‐structure segmentation directly on native diffusion images yielding moderate Dice scores between 0.63 and 0.80 compared to FreeSurfer T1w image segmentations applied to registered diffusion images [22]. However, the focus was not on extracting diffusion parameters from deep GM, but rather to improve anatomically constrained tractography of WM tracts. A similar deep learning approach used mean DWI and DTI parameter maps (MD and FA) generated from a 1‐mm isotropic high‐resolution acquisition of the hippocampus to train a segmentation model specifically for this structure, yielding accurate output relative to manual segmentation (Dice scores: 0.88–0.90) [23]. The third approach generated synthetic T1w images from diffusion data and then applied a standard T1w segmentation approach (e.g., FreeSurfer) to delineate various brain structures [24]. When compared to segmentations generated on the original T1w images, the segmentations based on the synthetic images showed significant overlap indicated by Dice scores ranging from 0.73 to 0.92 for separate deep GM structures. However, synthetically reconstructed T1w images also contain the same poor WM/GM contrast adjacent to deep GM structures that is apparent on native T1w images; hence, this method does not utilize a clear advantage DTI has over other imaging modalities for identifying the deep GM boundaries adjacent to the internal capsule. The fourth approach utilized deep learning to directly segment GM structures based on the raw diffusion signal itself [25] showing promising results compared to T1w segmentation (Dice scores ~0.75), suggesting that advantageous GM segmentation image contrast is present in the diffusion signal. However, even with these advances in deep learning segmentation techniques, the proposed approaches still retain the traditional limitations of deep learning–based segmentation in general: the difficulty in interpreting methods, the difficulty troubleshooting erroneous segmentations, and the unpredictable application to new datasets unused during training.

The purpose here is to develop an interpretable automated segmentation approach for deep GM that works solely on diffusion tensor images and maps that can be calculated from a large variety of typically acquired diffusion protocols. Like DDParcel [25], this technique avoids potential for erroneous registration with other scans like T1w, as well as not needing a T1w image in the first place as many clinical protocols do not acquire it. This is a natural extension of a previously developed surface‐based segmentation method for the cerebral cortex that also only relies on the contrast available from diffusion tensor images and maps [21]. The new method is evaluated on locally run 1.5‐mm isotropic whole‐brain diffusion acquisition data acquired in ~3 min in a test–retest cohort comparing regional brain volumes between the new method and FreeSurfer segmentations output from T1w images. As a first application, and yielding the first DTI healthy “lifespan” study across multiple deep GM structures, this novel method was used on 1.5‐mm isotropic diffusion MRI data to evaluate age changes of diffusion in the striatum, thalamus, globus pallidus, hippocampus, and amygdala in a cross‐sectional cohort of typical individuals (N = 357) across much of the lifespan (5–90 years).

2. Materials and Methods

2.1. Participants and Data Acquisition

The method described in Sections 2.2 and 2.3 was applied to a locally acquired test–retest high‐resolution diffusion dataset as well as a cohort of typical individuals spanning a large portion of the lifespan. The local healthy participant study was approved by the Human Research Ethics Board at the University of Alberta with written informed consent (and assent for children). Test–retest diffusion data were acquired from 24 healthy participants (29 ± 8 [20–48] years; 13 females) on a 3T Siemens Prisma (64‐channel head and neck coil) as part of the University of Alberta normative brain imaging database [26]. DTI was acquired with a single‐shot EPI spin‐echo sequence: multi‐band = 2, GRAPPA R = 2, 6/8 partial Fourier, 10 b0 s/mm², 6 b500 s/mm² (not used here), 20 b1000 s/mm², 64 b2500 s/mm² (not used here), TR = 5160 ms, TE = 67 ms, FOV = 220 mm, 96 1.5 mm slices with no gap, 1.5 × 1.5 mm², reconstructed over the 64 channels using sum of squares, with an acquisition time of 9:29 min (or 2:51 min for only the b0 and b1000 shells used for this study). For comparison to more standard T1‐based coregistered to DWI segmentation approaches, whole‐brain T1w imaging was acquired with an MPRAGE sequence: TE/TR/TI = 2.4/1800/900 ms, GRAPPA R = 3, 0.87 × 0.87 × 0.85 mm³ acquired resolution, 208 sagittal slices, and acquisition time of 3:39 min.

To assess age‐related DTI parameter trajectories in five deep GM structures, diffusion data were acquired from a separate lifespan cohort including 382 subjects recruited under the University of Alberta normative brain imaging database [26] and were scanned with the same 1.5‐mm isotropic diffusion imaging and T1w MPRAGE protocols as the test–retest cohort above. Amongst these individuals, 365 completed adequate diffusion and structural imaging with 357 (38 ± 23 [5–90] years; 203 females) remaining after subsequent quality control of subcortical segmentations (Section 3.4).

2.2. DTI Preprocessing

Preprocessing of the University of Alberta diffusion data (test–retest and lifespan cohorts) was performed with a custom pipeline designed such that only the diffusion images were used as follows. To generate consistent brain masks that were inclusive of all brain tissue voxels, initial tensor models were first fit on unprocessed diffusion data for each subject [27] (DIPY v1.0), and a brain mask was generated on the outputted MD map [28] (BET, FSL v6.0). These tensor maps were recalculated after the following preprocessing steps. Note that for the diffusion data, it was important to correct for the non‐central chi‐squared signal distribution resulting from the sum of squares image reconstruction of the 64‐channel phased array data, which was transformed to a Gaussian signal distribution using an automated method for estimating the non‐central chi‐squared distribution model parameters [29]. Diffusion data were then denoised using the non‐local spatial angular matching algorithm [30], corrected for Gibbs ringing [31], and corrected for motion and susceptibility induced/eddy current distortions [32] (eddy, FSL v6.0).

The final tensor models were fit (DIPY v1.0) outputting FA, MD, and primary eigenvector maps. Additionally, a mean b1000 DWI was output and was corrected for intensity bias using N4 correction based on the spatial variations of signal intensity estimated from the first b0 image [33]. A representative mean b1000 DWI is displayed along with FA and MD maps in Figure 1A–C. The mean b1000 DWI emphasizes the boundary of the globus pallidus, whereas the FA and MD maps clearly delineate the remaining boundaries of the striatum, thalamus, and hippocampus/amygdala region. In fact, WM and CSF voxels adjacent to deep GM can be classified accurately based on FA and MD alone [34]. A composite FA + MD map was created for each subject by adding the FA map to the MD map (in mm²/s, multiplied by 1000 to place MD values on a similar numerical scale to FA). The addition of these two maps was chosen because FA provides high contrast at GM/WM boundaries (WM has high FA, and GM has low FA), whereas MD provides high contrast at GM/CSF boundaries (CSF has high MD, and GM has lower MD) and virtually no contrast between GM and WM. By summing both maps, the composite image produces a single map where deep GM appears as a low‐intensity region surrounded by high‐intensity voxels at both the WM and CSF borders, creating strong image gradients at all relevant boundaries for the surface deformation algorithm (Figure 1D).

FIGURE 1.

(A) Mean b1000 DWI, (B) FA map, and (C) MD axial maps acquired with 1.5 × 1.5 × 1.5 mm³ resolution at 3T and (D) FA + MD (1000×) composite map in three orthogonal orientations. The mean b1000 DWI emphasizes the boundary of the globus pallidus (pink arrow). The FA map clearly delineates the boundary between deep gray matter structures and adjacent white matter, whereas the MD map emphasizes the boundary between the deep gray matter and CSF. When these maps are summed, the FA + MD composite map clearly delineates the boundaries of the striatum (green arrow), thalamus (yellow arrow), hippocampus (cyan arrow), and amygdala (red arrow).

2.3. Surface‐Based Deep GM Segmentation on DTI

Typical automated segmentation algorithms that utilize whole‐brain T1w images aim to segment seven separate subcortical GM structures in each hemisphere, namely, the putamen, caudate nucleus, nucleus accumbens, globus pallidus, thalamus, hippocampus, and amygdala. However, in our proposed method, the contrast on the FA + MD composite map is not sufficient to delineate the substructures of the striatum (nucleus accumbens, putamen, and caudate nucleus) and to separate the hippocampal head from the amygdala. Thus, our proposed method aims to segment five subcortical structures, namely, the striatum, globus pallidus, thalamus, hippocampus, and amygdala, where the striatum and hippocampus/amygdala regions are segmented based on the contrast available on the FA + MD composite map. To separate the hippocampus from the amygdala, a secondary segmentation step is performed based on expectation maximization segmentation (ANTs Atropos [35]). The entire diffusion‐only segmentation algorithm is as follows.

The Harvard–Oxford probabilistic atlas was registered to native DTI imaging space by calculating a nonlinear transformation between the FSL HCP FA template and the subject's FA map (ANTS SyN registration [36]) per participant extracting left/right probability maps for the caudate nucleus, putamen, globus pallidus, nucleus accumbens, thalamus, hippocampus, and amygdala. Initial voxel labels for each structure were generated by thresholding by a predetermined value to a singular cluster of voxels for each of the following four regions for the left and right hemispheres separately: (i) The globus pallidus label was created by thresholding the corresponding probability map by 50%; (ii) a striatum label was created by combining the caudate nucleus, nucleus accumbens and putamen probability maps and thresholding by 15%; (iii) a thalamus label was created using a threshold of 50%; and (iv) a hippocampus/amygdala region was created by thresholding hippocampus and amygdala probability maps by 35%. The different probability percentages were selected through a pragmatic visual optimization process across a subset of subjects spanning different ages, with the goal of producing a single connected cluster of voxels positioned close to the target boundary for each structure on most subjects. The lower threshold for the striatum (15%) is necessary because this label is created by combining three separate probability maps (caudate, putamen, and nucleus accumbens) whose spatial overlap is imperfect, requiring a more lenient threshold to produce a single continuous initial label. These thresholds do not critically affect the final segmentation because the subsequent surface deformation step refines the initial mesh to the true tissue boundaries visible on the diffusion images. However, for substantially different acquisition protocols or spatial resolutions, these initialization parameters may need to be re‐evaluated using the publicly available code. To remove any voxels containing primarily CSF or WM, voxels were removed from the labels with MD > 1.5 × 10⁻³ mm²/s or FA > 0.5, which are tensor parameter values outside of the range for GM [34]. Each label was then converted to a surface using the medical image registration toolkit (MIRTK) outputting 3D triangulated meshes consisting of vertices and edges.

Initial 3D meshes were then deformed (MIRTK, deform‐mesh) to image edges on the mean b1000 DWI for the globus pallidus and the FA + MD composite map for the remaining regions. Here, image edges are defined as spatial locations of maximum first‐order image intensity gradient (i.e., locations where the magnitude of the spatial derivative of image intensity is largest), indicating a transition between tissue types. The deformation algorithm iteratively moves each vertex of the 3D mesh toward the nearest image edge along the surface normal direction, whereas internal forces constrain surface smoothness. This external/internal force framework for deformable surface segmentation is well established in neuroimaging and forms the basis of widely adopted cortical segmentation tools such as FreeSurfer [16, 37], which deforms cortical meshes to tissue boundaries defined by image intensity gradients on T1w images. In the present method, this same framework is applied to deform subcortical meshes using the contrast available on diffusion images and maps, whereas internal forces regulate surface shape by restricting curvature and preventing self‐intersection [18, 20]. To accomplish native DTI deep GM segmentation, each structure is deformed separately using three categories of forces: (1) an external image‐based force that pulls each vertex toward the nearest image edge (maximum intensity gradient) on either the mean b1000 DWI or the FA + MD composite map depending on the structure; (2) an external atlas‐based force derived from the previously registered probabilistic atlas; and (3) internal forces that regulate surface shape by restricting the normal and Gaussian curvature of the surfaces. The atlas‐based force is calculated for each structure by summing the probability maps of adjacent GM structures and WM from the probabilistic atlas and subtracting the probability map of the structure being deformed. An example of the atlas‐based force is shown for the thalamus in Figure 2. This creates an inward force (positive values) in regions where neighboring structures are expected and an outward force (negative values) where the target structure is expected, constraining the surface deformation to image edges in the vicinity of the target structure. In regions where the tissue boundary is clearly defined on the diffusion images (e.g., the lateral border of the thalamus adjacent to the internal capsule), the image‐based force provides the primary guidance for boundary placement, whereas in regions where image contrast is weak (e.g., the inferior border of the thalamus), the atlas‐based force provides additional constraint. CSF probabilities were not included in the atlas‐based force because the contrast between deep GM and CSF on the FA + MD composite map is inherently strong due to the high MD of CSF, ensuring reliable image‐based boundary detection at GM/CSF interfaces without additional atlas constraint. Internal force parameters were set to the same value for each structure. Values and details for specific force parameters are provided in the accompanying code which is publicly available (https://github.com/LittleBrainLab/microbrain v1.0.0).

FIGURE 2.

Slice intersection of a 3D thalamus segmentation (yellow) generated using the proposed method shown on the (A) FA + MD composite map and (B) atlas‐based force generated from registered probabilistic atlas, where negative values (blue) indicate an outward facing force and positive values (red) indicate an inward facing force. Both maps are displayed as zoomed in with and without thalamus segmentation overlays. The segmentation algorithm utilizes both image‐based and atlas‐based forces; in the case of the thalamus, the image‐based force drives boundary placement where the boundary is clearly defined on the FA + MD map (superior and lateral regions), whereas the atlas‐based force provides additional constraint at the inferior edge of the structure where image contrast is weak.

The segmentation algorithm is visualized in Figure 3 and is summarized as follows. Initial segmentation models are used as input into the deformation segmentation method (Figure 3A) and are then deformed using the atlas‐based force and the image contrast on either the mean b1000 DWI image or the FA + MD composite map depending on the structure. The boundary between the globus pallidus and the putamen is clearly defined on the mean b1000 DWI given its lower signal intensity resulting from short T2*. Thus, to separate these structures, the initial surface model of the globus pallidus was first deformed to the closest edge on the mean b1000 DWI image (Figure 3B). The striatum was then deformed to the closest edge FA + MD composite map while restricting movement into the initial globus pallidus segmentation (Figure 3C). To prevent the left and right thalamus from intersecting along the medial boundary of these structures, both the left and right thalamus are deformed at the same time to the closest edge on the FA + MD composite map (Figure 3D).

FIGURE 3.

Segmentation workflow of deep GM structure based solely on diffusion tensor imaging visualized on diffusion images acquired with 1.5 × 1.5 × 1.5 mm³ resolution at 3T. (A) Initial surface models are created from registering a probabilistic atlas to native space. These structures are subsequently deformed one at a time using the mean b1000 diffusion weighted image (DWI) and the FA + MD composite map. (B) First, the globus pallidus (GP, pink) is deformed on the mean b1000 DWI. The remaining structures are deformed subsequently to the nearest edge on FA + MD composite map while restricting surface deformations into the previously segmented structures using the following order of striatum (C, green) and then thalamus (D, yellow). Given the lack of contrast between hippocampus and amygdala on DTI, these structures are segmented using multiple steps as follows. (E) First, the hippocampus/amygdala combined surface model is deformed to the nearest edge on the FA + MD map. Voxels within this region are then segmented into amygdala (red), hippocampus (cyan), and cortex voxels based on image intensity of the FA + MD map and registration to a probabilistic atlas (ANTS Atropos). Surface models are then created for each of these voxel‐based segmentations and deformed to the nearest edge on the FA + MD composite map. (F) Visualization of final deep GM segmentations for the globus pallidus, striatum, and thalamus (top and bottom) and for the left hemisphere for all five structures (left and right).

The lateral and medial boundaries of the hippocampus and amygdala are well defined on the FA + MD composite map; however, the boundary between the two structures is not easily discernible. Thus, the hippocampus and amygdala are segmented using multiple steps. First, the hippocampus/amygdala combined mesh is deformed to the closest edge on the FA + MD composite map (Figure 3E). Then, a voxel‐wise label of this region is extracted, and then these voxels are relabeled into hippocampus, amygdala, and cortex subregions with expectation maximization segmentation (ANTs Atropos [35]) using the Harvard–Oxford hippocampus, amygdala, and cortex probability maps previously registered to native diffusion imaging space via the nonlinear transformation described above. Surface meshes are then extracted for the hippocampus and amygdala regions. Then the hippocampus mesh is deformed again to the closest edge on the FA + MD composite map while restricting the surface deformation into the amygdala label, and the amygdala mesh is deformed to the closest edge on the FA + MD composite map while restricting the deformation into the hippocampus and cortex labels. These steps output 3D models of the hippocampus and amygdala where the lateral and inferior/superior boundaries of these structures are primarily determined by the contrast of the FA + MD composite map, and the boundary between the structures is primarily determined by the registration of the probability maps. For all subsequent analyses, the 3D surface meshes were converted to voxel‐wise labels by identifying the voxels enclosed within each mesh (Visualization Toolkit [VTK v9.2.6]). These voxel‐wise labels were used to calculate regional volumes (sum of enclosed voxels multiplied by voxel volume) and to extract mean FA and MD values (average across enclosed voxels).

2.4. Comparison of DTI‐Only Segmentation to Conventional Methods and Reproducibility of Proposed Method

2.4.1. T1‐Weighted Segmentation

To compare the segmentations generated from the proposed segmentation method based solely on DTI images and maps to a more conventional approach, deep GM segmentations were generated on the high‐resolution T1w images from the test–retest and lifespan cohort using FreeSurfer v7.4.1. Because the proposed diffusion‐only segmentation method segments the striatum as a single continuous structure, the T1‐based segmentations for the caudate nucleus, putamen, and nucleus accumbens were combined outputting a single striatum label for each subject in both cohorts. In a supplementary analysis, the HCP test–retest diffusion data and standard structural segmentation pipeline [38] was used as an additional comparison to the proposed technique. Given that the HCP diffusion data were corrected for spatial distortions, the influence of registration on the experimental design was reduced for this dataset relative to the experiment in the main manuscript that utilized distorted diffusion images. The following experiments were conducted on both the data acquired locally at the University of Alberta and the supplementary HCP test–retest data.

2.4.2. Segmentation Overlap Comparison

For each subject in the test–retest cohort (first scan only), the DTI‐based segmentations were registered using a 6° of freedom rigid transformation (FA map to T1 image [36]) and interpolated (nearest neighbors) to the T1 imaging space. The rigid registration was selected rather than a nonlinear registration as it provided a more accurate registration around regions of the internal capsule when registering the FA map to the T1 image. As a quantitative comparison between segmentation methods, Dice coefficients (overlap between labels) were calculated for each structure label from the proposed DTI segmentations (voxels contained within each 3D model) to the interpolated FreeSurfer T1w segmentations, outputting an inter‐method Dice coefficient per subject for the globus pallidus, striatum, thalamus, hippocampus, and amygdala, keeping left and right separate.

2.4.3. Structural Volume Comparison

In addition to Dice scores, regional brain volumes were calculated by summing the voxels enclosed within each 3D mesh (for the proposed method) or included in each FreeSurfer label (for the T1‐based method), multiplied by the respective voxel volume. Volumetric differences between the two methods were assessed statistically using paired t‐tests for each structure (p < 0.05) using the first scans of the test–retest data.

2.4.4. Segmentation Reproducibility Analysis

To enable a comparison between the Scan 1 and Scan 2 segmentations generated from the proposed method, a rigid affine transformation was calculated from the Scan 2 mean b0 image to the Scan 1 mean b0 image for each subject [39] in the test–retest dataset. The transformation was then applied to the segmentations from Scan 2 and nearest neighbors interpolated to native imaging space on the diffusion images of Scan 1. Inter‐scan Dice coefficients were then calculated between the segmentations from Scan 1 and Scan 2. Additionally, intraclass correlation coefficients (ICC(3,1), two‐way mixed effects, single measures, absolute agreement) with 95% confidence intervals were calculated between Scan 1 and Scan 2 volumes for each structure using the Python pingouin package (v0.6 [40]).

2.5. Lifespan Deep GM Diffusion Measurements

The proposed DTI segmentation method was performed on the diffusion data from the lifespan imaging cohort for each subject. 3D mesh segmentations were overlaid on the mean b1000 DWI and FA + MD composite map for each subject to qualitatively assess accuracy of the subcortical segmentations. For each subject, MD and FA were calculated as the average from voxels enclosed within each of the 3D mesh structures and volumes were calculated by summing the enclosed voxels of each structure. To compare volumetric trajectories to those from conventional T1‐based methods, volumes were generated using FreeSurfer v7.4.1 [41] on the lifespan cohort. Hemispheric differences were assessed for each structure (pairwise t‐test, p < 0.05), and because no interhemispheric differences were detected for FA or MD, the values were averaged between hemispheres. Average MD and FA values were then calculated for each structure (left/right averaged) separately. To assess changes in volume and diffusion parameters related to age and sex, regression modeling was conducted for the lifespan cohort as follows. FA, MD, volume (DTI segmentation), and volume (FreeSurfer T1‐based) trajectories were estimated separately for each structure with linear, quadratic, and cubic curves using ordinary least squares regression implemented in the Python statsmodels package (v0.14, OLS routine) [42]; significant curves (p < 0.05) with the lowest Akaike Information Criterion (AIC) values were considered the best fit. Paired and independent samples t‐tests were performed using the SciPy stats module (v1.16, scipy.stats.ttest_rel, and scipy.stats.ttest_ind) [43]. In a supplementary analysis, average FA, MD, and volume (DTI segmentation) were calculated separately for males and females for each structure and an independent samples t‐test (p < 0.001 uncorrected for age and multiple comparisons) was used to assess statistical differences associated with sex. Subsequently, the same model fitting procedure was applied separately for males and females to investigate age‐related trajectories for each measure by sex.

3. Results

3.1. Assessment of Deep GM Segmentations on Native DTI

Based on qualitative assessment for the 1.5‐mm isotropic University of Alberta test–retest diffusion data, the segmentation algorithm performed adequately with segmentations of the globus pallidus, striatum, and thalamus (example Figure 4) for all 24 subjects bordering the high‐contrast regions visible in the diffusion images. However, some instances of incorrect segmentations were observed for the hippocampus and amygdala regions. In these data, the right hippocampus segmentation expanded laterally into the cortex on two subjects, with the same segmentation error observed in the bilateral hippocampus segmentations of a separate subject. Amygdala segmentations consistently delineated GM voxels; however, the anterior boundary of this segmentation often included cortex due to the lack of contrast between amygdala and cortex on the FA + MD composite map.

FIGURE 4.

Example segmentation output generated for a 33‐year‐old, male participant from the test–retest dataset used for the validation of the proposed DTI‐only segmentation algorithm: Segmentations are displayed for the globus pallidus (pink), striatum (green), thalamus (yellow), hippocampus (cyan), and amygdala (red). Visualizations of slice intersections of deep GM surface models are displayed on (A) the mean b1000 diffusion weighted image and (C) FA + MD composite map, as well as (B) a 3D rendering of the surface models.

3.2. Comparison to Standard T1‐Weighted Segmentation Method

An example of the head‐to‐head comparison of the T1‐based (0.7 mm isotropic) FreeSurfer segmentation pipeline and the proposed diffusion‐only method is displayed in Figure 5. Although the majority of the segmentations were in agreement, FreeSurfer produced errors consistently in a few regions that are clearly observed when viewed on the rigidly registered FA map (FA to T1w). FreeSurfer consistently included voxels in the globus pallidus and thalamus segmentations belonging to the internal capsule (Figure 5B,E). FreeSurfer also consistently included WM voxels along the lateral portion of the striatum segmentation adjacent to the putamen (Figure 5B,E). Importantly, these segmentation errors were already present when the FreeSurfer output was viewed in its native T1w imaging space (Figure 5A,D), confirming that they originate from insufficient GM/WM contrast on the T1w images rather than from any misalignment introduced by the rigid registration of the FA map to the T1w space used for visualization. In contrast, the DTI‐based deep GM segmentation method properly excluded adjacent WM voxels for these structures primarily because of the reliance on the WM/GM contrast present on the FA map (Figure 5C,F). A supplementary analysis of the distortion corrected test–retest HCP data (T1w 0.7 mm isotropic resolution, HCP 1.25 mm isotropic resolution) is visualized in Figure S1, demonstrating similar errors relative to the undistorted diffusion data and further supporting the notion that the poor T1w GM/WM contrast is driving the segmentation errors adjacent to the internal and external capsules rather than problematic registration.

FIGURE 5.

Voxel‐wise segmentation generated from FreeSurfer on the two subjects from the test–retest data (top row: 37 year old, female, bottom row: 22 year old, female) overlaid on an axial slice from a T1w structural image (A,D) and the subject's fractional anisotropy map rigidly registered to the T1 image (B,E). Slice intersections of 3D models generated from the proposed DTI‐only segmentation method overlaid on the FA map in diffusion imaging space from the approximately same slice relative to the T1w image in the same subject (C, F). Segmentations are displayed for the globus pallidus (pink), striatum (green), thalamus (yellow). In these examples, the voxel‐wise FreeSurfer segmentation poorly delineates the edge of these structures reflecting the poor T1w GM/WM contrast (second panels in A and D) in regions adjacent to these structures.

Group average (test–retest first scan) subcortical GM volumes for the left/right hemispheres for all five structures are presented for each method and for the test–retest (first scan only) comparison to FreeSurfer T1 segmentation (Table 1). Relative to the standard T1w segmentation algorithm (FreeSurfer), the proposed DTI‐only segmentations generated statistically smaller volumes (percent difference between −9% and −38%) for all 10 structures, likely reflecting the lower spatial resolution of the diffusion data along with the aforementioned tendency of T1w GM segmentation to include WM voxels. Even with statistically different volumetric measurements, a large proportion of the voxels were in agreement between the two approaches, as evidenced by Dice scores larger than 0.70 in 5 of the 10 structures. The left/right thalamus had Dice scores greater than 0.80, whereas the left/right globus pallidus, left/right amygdala, and right hippocampus segmentations had Dice scores lower than 0.70 (Table 1). In supplementary analysis of the higher resolution HCP test–retest data, similar but smaller volumetric differences were observed with Dice scores for each structure in the range of the values observed in the University of Alberta data (Table S1).

TABLE 1.

Group average volume measurements for the test–retest cohort extracted per structure using standard 3D T1‐weighted automated segmentation (FreeSurfer) and the proposed diffusion‐only segmentation method (Scan 1 only).

Test–retest data (n = 24, Scan 1 only)
	3D T1‐weighted FreeSurfer (volume mm3)	Proposed diffusion‐only method (volume mm3)	T1w vs. DTI‐only percent volume difference	Volume difference (p)	Inter‐method Dice
Left striatum	8493 ± 728	7751 ± 540	−9%	< 0.001	0.73 ± 0.04
Right striatum	9029 ± 666	7904 ± 487	−12%	< 0.001	0.74 ± 0.03
Left globus pallidus	2182 ± 224	1347 ± 88	−38%	< 0.001	0.64 ± 0.10
Right globus pallidus	2184 ± 268	1372 ± 119	−37%	< 0.001	0.69 ± 0.11
Left thalamus	8875 ± 718	6614 ± 509	−25%	< 0.001	0.80 ± 0.03
Right thalamus	8315 ± 790	6954 ± 631	−16%	< 0.001	0.84 ± 0.02
Left hippocampus	4481 ± 370	3323 ± 270	−26%	< 0.001	0.70 ± 0.03
Right hippocampus	4600 ± 364	3256 ± 370	−29%	< 0.001	0.69 ± 0.04
Left amygdala	1559 ± 185	1281 ± 237	−18%	< 0.001	0.50 ± 0.08
Right amygdala	1761 ± 141	1499 ± 188	−15%	< 0.001	0.59 ± 0.07

3.3. Reproducibility Analysis

Using the test–retest cohort, volumes and test–retest Dice scores were calculated for all five subcortical structures keeping left and right hemispheres separate (Table 2). The proposed DTI‐only segmentation method was highly reproducible indicated by high Dice scores ranging from 0.79 (left amygdala) to 0.94 (left/right thalamus) and high ICC values for absolute agreement ranging from 0.67 (right amygdala) to 0.98 (right thalamus) (Table 2). ICC values were highest for larger structures (striatum: 0.92–0.97, thalamus: 0.96–0.98) and lower for smaller structures (globus pallidus: 0.80–0.83, amygdala: 0.67–0.75). The lower ICC values for smaller structures likely reflect the disproportionate effect that small absolute voxel‐level variability has on the ICC when the total structure volume is small, rather than a meaningful reduction in segmentation reliability. Paired t‐tests revealed no statistical difference (all p > 0.05 uncorrected) in volume measurements between Scan 1 and Scan 2 for all structures. Notably, in the retest data, the same segmentation errors noted for the hippocampus and amygdala on the Scan 1 data were reproduced in the exact same regions and subjects for the corresponding data at Scan 2. In the supplementary HCP analysis, similarly high inter‐scan Dice scores were observed for all structures (ranging from 0.88 to 0.93) (Table S2).

TABLE 2.

Reproducibility of proposed diffusion‐only segmentation method over two scans. Group average volumetric measurements for the test–retest cohort extracted per structure using the proposed diffusion‐only segmentation method along with inter‐scan dice coefficients and intraclass correlation coefficients (ICC) along with 95% confidence intervals (CI). No significant pairwise differences (p < 0.05) were observed between Scan 1 and Scan 2.

University of Alberta test–retest (n = 24)
	Scan 1 volume [mm3]	Scan 2 volume [mm3]	Inter‐scan Dice	ICC (95% CI)
Left striatum	7751 ± 540	7691 ± 578	0.91 ± 0.02	0.97 (0.92–0.98)
Right striatum	7904 ± 487	7869 ± 578	0.91 ± 0.02	0.92 (0.82–0.96)
Left globus pallidus	1347 ± 88	1344 ± 112	0.89 ± 0.03	0.83 (0.65–0.92)
Right globus pallidus	1372 ± 119	1378 ± 108	0.89 ± 0.04	0.80 (0.58–0.91)
Left thalamus	6614 ± 509	6570 ± 597	0.94 ± 0.02	0.96 (0.90–0.98)
Right thalamus	6954 ± 631	6916 ± 622	0.94 ± 0.02	0.98 (0.95–0.99)
Left hippocampus	3323 ± 270	3283 ± 361	0.87 ± 0.03	0.84 (0.67–0.93)
Right hippocampus	3256 ± 370	3259 ± 423	0.88 ± 0.03	0.92 (0.83–0.97)
Left amygdala	1281 ± 237	1275 ± 212	0.79 ± 0.07	0.75 (0.51–0.88)
Right amygdala	1499 ± 188	1441 ± 180	0.84 ± 0.05	0.67 (0.38–0.84)

3.4. Deep GM DTI Metrics and Volume Versus Age

The diffusion‐only segmentations worked well across the lifespan as shown for three subjects with very different ages (Figure 6). Notably, in regions with lost signal due to T2* decay observable in older subjects on the mean b1000 DWI, segmentations of the globus pallidus and striatum were also reconstructed accurately (example Figure 6, 82 year old subject). Amongst all subjects in the lifespan dataset, five were excluded because of erroneous amygdala segmentations caused by the poor signal around the air tissue interface adjacent to the inferior frontal region of the brain. FreeSurfer failed to reconstruct segmentations on T1 images for three subjects, which were excluded from further analysis. After excluding subjects, N = 357 (37 ± 22 [5–90] years; 203 females) remained for modeling of lifespan trajectories.

FIGURE 6.

Segmentations generated from the proposed DTI‐only method displayed for three participants spanning a large age range in the lifespan cohort. Segmentations are displayed in 3D (row 1) as well as with and without segmentation overlays for the mean b1000 DWI image (rows 2–4) and the FA + MD composite map (rows 5–7). All 5 regions output from the proposed method are displayed including the globus pallidus (pink), striatum (green), thalamus (yellow), hippocampus (cyan) and amygdala (red). Even with substantial subject variability in brain shape, the segmentation algorithm generates structures that align well with the image edges visible on the diffusion images.

Across all subjects, distinct FA values were observed for each subcortical GM structure with higher FA values being observed for the thalamus (0.25 ± 0.01) and globus pallidus (0.23 ± 0.02) compared to the amygdala (0.18 ± 0.02), striatum (0.16 ± 0.01), and hippocampus (0.14 ± 0.01). MD values were lower for the globus pallidus (0.78 ± 0.04 × 10⁻³ mm²/s) and striatum (0.81 ± 0.02 × 10⁻³ mm²/s) as compared to the amygdala (0.86 ± 0.02 × 10⁻³ mm²/s), thalamus (0.87 ± 0.03 × 10⁻³ mm²/s), and hippocampus (0.93 ± 0.03 × 10⁻³ mm²/s).

All five deep GM regions showed significant age‐related changes in FA and/or MD over 5–90 years (Figure 7), with the shape of the trajectory varying by structure and metric (8 nonlinear fits, 2 linear fits; all p < 0.05). The FA age trajectories differed per region, with globus pallidus and hippocampus yielding quadratic fits that were flat from ages 5 to ~60 years of age and then became smaller, amygdala with negative linear correlation, striatum with positive linear correlation, and thalamus with little change (relatively stable cubic trajectory across age). MD values for all structures showed quadratic or cubic U‐shaped fits with minima at ~30–40 years, where notably the drop of MD during development to the minima was not as much as the greater MD with aging into the elderly, particularly for the thalamus and hippocampus.

FIGURE 7.

FA (row 1) and MD (row 2) values (left/right averaged) for all subjects from the lifespan cohort (N = 357; 5–90 years) plotted by age for the globus pallidus (pink), striatum (green), thalamus (yellow), hippocampus (cyan) and amygdala (red). Scatter plots in those colors are displayed with the best fitting model (black ‐ chosen from linear, exponential, quadratic and cubic fits). FA values remained flat until age 5 to ~25 years old for the globus pallidus, thalamus and hippocampus but increased linearly for the striatum and decreased linearly for the amygdala. MD values decreased from childhood to early adulthood (~30–35 years) and increased thereafter resulting in higher MD values in the elderly compared to childhood.

Different developmental and aging volume trajectories were observed for each structure with marked differences depending on T1 versus DTI segmentation (Figure 8). In the lifespan dataset, while using the T1‐based FreeSurfer segmentation (Figure 8, Column 1), a decreasing linear trajectory was observed for the globus pallidus, and a decreasing quadratic trajectory was observed for the striatum. The thalamus and hippocampus had volume trajectories that were flat from 5 to ~35 years and then declined thereafter, whereas little change was observed through 5–90 years for the amygdala. Similar volumetric trajectories were observed when using the DTI‐only segmentation method for the striatum (linearly decreasing), thalamus (cubic flat from 5 to ~35 years then decreases), and amygdala (flat quadratic trajectory), whereas the volumetric developmental and aging trajectory of the hippocampus increased from 5 to ~40 years and then decreased contrary to the T1 approach (flat from 5 to ~35 years). The biggest difference between developmental and aging trajectories was in the childhood/adolescent age range of the volumetric trajectory for the globus pallidus. With the DTI‐only method, a cubic trajectory was observed that increased from 5 to 30 years and then leveled off and decreased thereafter, contrary to the T1 approach (decreased from 5 to ~25 years).

FIGURE 8.

Volume (left/right averaged) calculated on high resolution T1‐weighted image using FreeSurfer (Column 1) and volume calculated using the proposed DTI‐only segmentation method (column 2) for all subjects from the lifespan cohort (N = 357, black curve fits) plotted by age for the globus pallidus (pink), striatum (green), thalamus (yellow), hippocampus (cyan) and amygdala (red). Notable age‐related volume differences are apparent between methods (T1‐based segmentation volume minus DTI‐only segmentation volume in column 3). In particular, the globus pallidus T1 segmentation suggests a decrease in volume from 5 to 30 years whereas DTI segmentation suggests an increase over the same age range. The volumes from DTI segmentation were smaller than T1 volumes across the entire age range except for the amygdala.

Average volumes were higher for the T1‐based segmentation compared to the DTI‐only volume across subjects for all five deep GM structures with the following volume differences (T1w volume–DTI volume): globus pallidus (858 ± 196 mm³), striatum (1013 ± 476 mm³), thalamus (1744 ± 426 mm³), hippocampus (1148 ± 333 mm³), and amygdala (211 ± 258 mm³). The T1 segmentation versus DTI segmentation volume differences differed as a function of age (Figure 8, Column 3). The only case with minimal volume differences between methods was the amygdala (Figure 8O) with the volume differential between methods lessening with age. At younger ages, the globus pallidus showed a marked reduction in this volumetric T1w–DTI segmentation mismatch from 5 to 20 years, and then the mismatch leveled off with aging into the elderly. The striatum showed a U‐shaped T1w versus DTI volume difference with age. The T1 versus DTI volume difference was smaller with age (although still sizeable) for the hippocampus.

3.5. Sex Related Differences in Deep GM DTI Metrics

Using the proposed DTI‐only segmentation method, differences associated with sex were assessed for FA, MD, and volume using the lifespan dataset (Figure S2). Developmental and aging trajectories were similar between males and females for FA and MD for all five structures, whereas volumetric trajectories differed slightly between sexes for several structures (Figure S2). No sex differences were observed for FA (left/right averaged) of the five deep GM structures (Figure S2, Column 1). MD (left/right averaged) was higher (p < 0.001 uncorrected) in males compared to females for four out of the five structures (globus pallidus [males: 0.79 ± 0.04 × 10⁻³ mm²/s, females: 0.77 ± 0.04 × 10⁻³ mm²/s; p < 0.001], striatum [males: 0.82 ± 0.03 × 10⁻³ mm²/s, females: 0.81 ± 0.02 × 10⁻³ mm²/s; p < 0.001], thalamus [males: 0.88 ± 0.03 × 10⁻³ mm²/s, females: 0.87 ± 0.03 × 10⁻³ mm²/s; p < 0.001], and hippocampus [males: 0.93 ± 0.03 × 10⁻³ mm²/s, females: 0.92 ± 0.03 × 10⁻³ mm²/s; p < 0.001]), whereas no sex difference was observed for the amygdala. All five structures (left/right average) had larger volumes in males compared to females: globus pallidus (males: 1346 ± 181 mm³, females: 1230 ± 147 mm³; p < 0.001), striatum (males: 8072 ± 1048 mm³, females: 7464 ± 1084 mm³; p < 0.001), thalamus (males: 6723 ± 798 mm³, females: 6360 ± 773 mm³; p < 0.001), hippocampus (males: 3369 ± 357 mm³, females: 3097 ± 377 mm³; p < 0.001), and amygdala (males: 1521 ± 234 mm³, females: 1368 ± 200 mm³; p < 0.001). Development and aging FA and MD trajectories were similar between males and females for all five structures, whereas qualitative differences in volumetric trajectories were observed for four of the five structures when comparing males and females on the sex‐stratified plots (Figure S2). Specifically, the globus pallidus, thalamus, and amygdala appeared to have the same volume for males and females at 5 years of age but increased in males toward 30–50 years of age, and the male striatum reduced in volume at a slower rate relative to females.

4. Discussion

4.1. Segmentation on DTI Alone

A subcortical GM segmentation method is proposed that uses the contrast available on DTI maps and the mean b1000 DWI to segment the globus pallidus, striatum, thalamus, hippocampus, and amygdala, removing the need for the acquisition of or registration to a T1w structural image. Because the segmentation primarily relies on the contrast of the mean DWI and the proposed quantitative FA + MD (1000×) composite map, it can be applied on various high‐resolution DTI acquisitions because FA and MD values for GM, WM, and CSF are generally similar between varying protocols. Qualitatively, the proposed segmentation approach performed similarly when applied to the test–retest cohort and the University of Alberta Normative database, both of which used 1.5‐mm isotropic diffusion data and spanned a wide age range. The method has been combined with a previously proposed DTI‐only cortex segmentation method [21] and is available as part of a fully automated diffusion MRI post processing and GM segmentation pipeline (https://github.com/LittleBrainLab/microbrain v1.0.0).

4.2. Comparison of DTI‐Only Segmentation to Standard Approaches

Compared to standard T1‐based FreeSurfer segmentation, the proposed DTI segmentation method was largely in agreement, indicated by high inter‐method Dice scores for the striatum (left: 0.73, right: 0.74), thalamus (left: 0.80, right: 0.84), and left hippocampus (0.70) but lower for the globus pallidus (left: 0.64, right: 0.69), amygdala (left: 0.50, right: 0.59), and right hippocampus (0.69) segmentations. Three previous studies have proposed DTI‐only subcortical GM segmentation using machine learning techniques. In a study proposing DORIS, a deep learning–based segmentation algorithm, similar FreeSurfer segmentation agreement was reported compared to the current study (globus pallidus: 0.66, striatum [consisting of putamen: 0.78 and the caudate nucleus: 0.63, nucleus accumbens excluded], thalamus: 0.78, hippocampus: 0.80 and amygdala: 0.68) [22]. Another approach (deepAnat) uses deep learning to create synthetic T1w images from diffusion images and inputs these synthetic images into FreeSurfer. Higher FreeSurfer segmentation overlap was observed using deepAnat (globus pallidus: 0.82, striatum [consisting of the putamen: 0.88, caudate nucleus: 0.88 and nucleus accumbens: 0.74], thalamus: 0.91, hippocampus: 0.88 and amygdala 0.82) [24] compared to the proposed method of the current study. The higher deepAnat Dice scores compared to the proposed method are not surprising given that deepAnat aims to recreate the same T1w contrast used by FreeSurfer. However, as shown in Figure 5, the T1w contrast may not be ideal for segmenting subcortical GM especially in regions around the internal/external capsules, whereas the DTI maps may provide advantageous contrast for delineating these structures (Figure 1). The final deep learning approach, DDParcel [25], is most similar to the proposed segmentation technique, training a deep learning segmentation algorithm directly on the diffusion data but using registered T1w‐based segmentation as training input. This results in on average higher Dice scores than previous deep learning–based approaches (Dice scores ~0.75 across all cortical and subcortical structures), supporting the idea put forward in the current study that the diffusion contrast is valuable for segmenting GM. Because the proposed method uses an interpretable segmentation algorithm, one clear advantage of the proposed method over deep learning–based approaches is that it can be intuitively “debugged” when errors arise. Thus, future improvements can be made to the technique without requiring retraining and/or large datasets. Importantly, although all methods here use FreeSurfer output as a “gold standard” segmentation comparison, FreeSurfer itself can generate errors visible when segmentations are overlaid on DTI maps, which is an alternative explanation to imperfect segmentation Dice scores reported with all previous DTI‐based segmentation methods. The proposed method was intentionally developed independent of other segmentation approaches, whereas previous deep learning–based approaches use FreeSurfer segmentation output for model training and testing (e.g., [22, 25, 44]) or use synthetic T1w images as input into FreeSurfer [24]; thus, these techniques may be capable of generating the same segmentation errors generated by FreeSurfer.

In the test–retest data, higher volumes were measured with FreeSurfer compared to the proposed method for all five subcortical GM structures. The larger T1‐based FreeSurfer volumes compared to the proposed method could be explained by many factors, two of which are (1) imaging resolution of the T1w imaging, which is far superior to that of the diffusion images causing thinner regions and boundary voxels to be included for a given structure, and (2) the poor T1w image contrast, which causes labels to reach past their true extent, a problem that is mitigated by the contrast available in the diffusion images and maps. Future work will aim to better understand the impact of the diffusion/T1w contrasts on image segmentation with the aim of matching imaging resolution between modalities, a notably challenging problem given the difference in noise levels and effective resolution (among many others) between T1w and diffusion imaging techniques.

4.3. Deep GM Developmental and Aging DTI Trajectories

Few studies have examined DTI lifespan (excluding the first 5 years) trajectories of the five deep GM structures targeted in this study. The proposed DTI‐only segmentation method enabled this analysis on 357 locally acquired subjects spanning 5–90 years. MD and FA values extracted per structure across age were in the expected range for GM (FA: ~0.15 to ~0.3, MD: ~0.75 to ~1.0 × 10⁻³ mm²/s) and differed across structures, suggesting that the diffusion tensor is sensitive to changes in microstructure. Relative to the DTI development and age trajectories of WM across the lifespan [45, 46] (for review, see [47]), subcortical GM trajectories were more subtle across 5–90 years. In general, MD versus age yielded a U‐shaped trajectory, which became smaller until mid‐adulthood and then increased thereafter for all structures. The MD minima of ~20 to ~40 years is comparable to the MD minimum observed for WM tracts across the lifespan [45]. In GM, decreases in MD during the first third of life may reflect one of many developmental processes, resulting in increased barriers to diffusion (e.g., dendritic arborization, axonal ramification, synaptogenesis, and glial proliferation), whereas the subsequent increases in MD during aging may be associated with the loss of such barriers (e.g., cell loss). FA trajectories showed some increases during development for 4/5 structures (except amygdala) in line with what is seen for WM tracts, albeit by a lot less, but the subsequent decrease with aging to yield an inverted U trajectory was only observable for globus pallidus and hippocampus. The amygdala shows a steady linear decrease in FA, but given the myriad of cells composing this structure [48], a physiological cause of such a trajectory is unclear. In contrast, the FA of the striatum increased with age over the entire lifespan, in agreement with other aging studies reporting linear increases in FA for the putamen (part of the striatum) [10, 12, 49], but an observation not seen in WM [45]. During healthy aging, the striatum increases in iron accumulation, resulting in faster T2* decay [50] and a lower diffusion MRI SNR in this structure. Lower SNR resulting from faster T2* results in the measurement of a higher FA and marginally lower MD [51], a well‐known effect when measuring the diffusion tensor in low‐SNR environments [52]. Thus, the linear developmental aging FA trajectory observed for the striatum may reflect an increase in iron accumulation rather than other underlying microstructural change. To avoid the SNR aging confound in the future, the proposed segmentation method could be leveraged to automatically exclude regions of the striatum with this low signal intensity measured on the mean b1000 DWI.

4.4. Developmental Alterations in Deep GM DTI

Developmental changes measured by DTI for multiple deep GM structures have been reported previously in studies using manually defined ROIs but excluded the hippocampus and amygdala [5, 8, 10]. In general, increasing FA and decreasing MD were observed over childhood and adolescence for varying basal ganglia structures (putamen, caudate nucleus, globus pallidus, and thalamus) [5, 8, 10] with rates of change similar to trajectories from the current study for the striatum. The human striatum primarily consists of medium spiny projection neurons and their dendritic spines [53]; thus, in development, in addition to iron accumulation, an increase in FA and decrease in MD may reflect an increasingly organized development of dendritic spines throughout adolescence. The increases in FA observed for the globus pallidus in previous neurodevelopmental studies [5, 8, 10] were not evident in the age trajectories in the current study, potentially reflecting the large advances in diffusion imaging/analysis (e.g., higher resolution, higher SNR and measurements extracted from the entire globus pallidus structure) compared to earlier work. Specifically, the lower resolution of previous studies likely suffered from greater partial volume contamination from the adjacent highly anisotropic internal capsule, possibly artificially driving FA increases that are minimized by the current high‐resolution surface‐based segmentation. Interestingly, a decrease in MD was observed for the globus pallidus over development in the lifespan dataset, suggesting that the changing DTI measures over development reflect an increase to the hindrance of diffusion rather than an increasingly organized microstructure. No developmental change in MD or FA was observed for the thalamus, suggesting little to no microstructural maturation of this structure in development. This stability contrasts with earlier reports of decreasing MD developmentally using a manually delineated subregion of the thalamus [5, 8]. This discrepancy may arise from the averaging of diffusion metrics across the entire heterogeneous thalamus in the current whole‐structure segmentation masking subtle regional maturation, or arguably could reflect the lower spatial resolution and signal‐to‐noise ratio of previous data compared to the high‐resolution acquisition used here. The amygdala and hippocampus were not measured in earlier 1.5‐T DTI deep GM development studies with limited spatial resolution (3‐mm‐thick slices) [5, 8, 10], but their measurement was enabled by the proposed segmentation method and the higher resolution diffusion acquisition used in the current study. Hippocampus MD got smaller over development in the current study, whereas FA was unchanged over development. The FA and MD developmental trajectories were similar to those measured from a high‐resolution 1‐mm isotropic diffusion acquisition specifically designed for the hippocampus and acquired on two lifespan cohorts; the same cohort as the current study [23] and a separate lifespan cohort of 153 participants [54]. The high‐resolution developmental MD and FA trajectories matched those measured in the current study supporting the notion of increased FA in development, which may relate to increased axonal packing and/or myelination [55, 56]. Mean diffusivity has been observed to decrease for the amygdala plateauing at ~10 years [6] earlier than the MD plateau observed in the current study into young adulthood. A shallow linear decreasing trajectory was observed across development for FA of the amygdala, suggesting that the DTI changes observed over this time period are not related to changes along a single axis of the diffusion tensor, whereas previous work has suggested that diffusion along the primary diffusion tensor direction is associated to fibers passing through the amygdala [57]. Taken together, these studies suggest that DTI is sensitive to continued development of these structures throughout childhood and into late adolescence toward young adulthood, contrary to previous work suggesting deep GM FA and MD trajectories level off prior to 4 years old [58].

4.5. Aging Related Changes in Deep GM DTI

In middle adulthood, MD trajectories modeled for the lifespan cohort from the current study for all five structures level off and begin to increase into later adulthood. The inversion of deep GM MD trajectories was not reported in previous studies that investigated younger cohorts (max age of 40 years [4] and 52 years [10]), an advantage of the current studies' larger age span. Mid‐life FA trajectories declined (globus pallidus, hippocampus, and amygdala) or plateaued (thalamus) in all structures but the striatum, which showed a steady linear increase throughout middle adulthood. Increases in MD trajectories were observed after ~30–35 years in all five deep GM structures with the most prevalent increases observed in the thalamus and hippocampus. Importantly, the entire medial boundaries of these two structures are adjacent to CSF; thus, age‐related atrophy of these structures may increase the proportion of partial volume measurement of CSF, hence increasing MD. In a high‐resolution (1 mm isotropic) DTI lifespan studies of the hippocampus, no significant increase in MD was observed in adulthood in two cohorts; a separate lifespan cohort of 153 participants [54] and the same cohort used in the current study [23]. However, the manual segmentation protocol [54] and deep learning–based segmentation [23] used in these previous studies did not include the subiculum a region bordering CSF over the anterior–posterior extent of the hippocampus, suggesting that the increase in MD in the current studies' hippocampus trajectories was at least partially driven by regions prone to CSF contamination. Interestingly, even the globus pallidus, a region encapsulated by surrounding WM, and the putamen also showed an MD increase in adulthood. Previous findings have observed larger globus pallidus MD values in adult populations compared to young adults using standard DTI acquisitions [11] and CSF suppressed FLAIR DTI [9], suggesting that the observed increase in MD observed in the current study is not solely due to CSF partial volume measurement.

Decreasing or constant FA trajectories were observed into adulthood of four of the five deep GM structures with the exception of the striatum that had a steady linear increase in FA throughout life. Aging‐associated FA increases in the striatum (putamen and/or caudate nucleus) have been consistently reported in both lifespan studies [10, 12, 59, 60] and with higher FA in older subjects reported in cross‐sectional comparisons between old and young cohorts [9, 11]. This increase in FA has been linked to increasing iron accumulation in this structure [51] with FA increases in the striatum into adulthood also being observed in the current study. Our replication of this linear striatal FA increase provides further evidence that this specific deep GM aging trajectory is robust across different cohorts and acquisition protocols. Another explanation of increasing FA with age in the striatum may be the targeted loss of neurons or particular dendrite connections [59]. The remaining four structures either had no change into adulthood (thalamus) or decreased into late adulthood (globus pallidus, hippocampus, and amygdala). Decreases in FA were not observed in previous studies that did not contain elderly participants (max age of 40 years [4] and max age of 52 years [10]). In the high‐resolution hippocampus lifespan studies previously mentioned [23, 54], decreased FA was also observed, indicating that this age‐related change is likely not related to partial volume effects of the hippocampus with CSF. A variety of factors may explain decreases in FA in these three structures; one such explanation is the demyelination of fibers that traverse all three of these structures (e.g., strato‐palidonigral bundle in the globus pallidus, the fimbria/alveus of the hippocampus, and other projection fibers such as the fornix passing through the amygdala [57]). However, other microstructural alterations (e.g., neuronal loss and decreased axon packing) may also alter tensor parameters as a function of age and thus cannot be ruled out.

4.6. Development and Aging Volume Trajectories of Deep GM Measured on DTI

Development and aging volume trajectories for four (striatum, thalamus, hippocampus, and amygdala) of the five subcortical GM structures investigated were largely similar when using the proposed DTI‐only segmentation method on the DTI images compared to the ubiquitous T1‐based volumes (FreeSurfer segmentation in this case). However, the T1 segmentations yielded consistently larger volumes than the DTI segmentations, as previously discussed. This means that the developmental and aging volumetric interpretation would remain the same regardless of using the proposed segmentation method or FreeSurfer with these data. Notably, much of the lifespan volumetric trajectories agreed between the proposed method and FreeSurfer, for example, the large reduction in thalamus volume observed throughout late adulthood and the continued decrease in striatum volume observed across the entire lifespan (5–90 years). In contrast to previous reports of earlier adolescent peaks [1, 3, 61], the volume trajectories reported in this study for the globus pallidus, thalamus, and hippocampus indicate a shallow growth phase with later development peaking or plateauing around age 30. This discrepancy likely stems from our use of quadratic and cubic modeling, as opposed to the exponential [1, 61] or Poisson fitting [3] techniques used in prior literature that can mathematically force earlier inflection points during the transition to adolescence. Furthermore, the continuous linear volumetric decrease observed here for the striatum differs from high‐resolution T1 studies that have shown distinct adolescent peaks for its substructures, such as the putamen (~12 years) and caudate nucleus (~14 years) [3, 61]. Given the large natural range of striatal volume during development, these conflicting findings may reflect differences in the specific populations sampled across different lifespan studies rather than a definitive lack of a developmental growth phase.

The globus pallidus prior to ~20 years had a decreasing volume when measured with FreeSurfer on T1w relative to the DTI‐only method that had an increasing volume during this age range. It is unclear whether the globus pallidus increases or decreases in volume during childhood/adolescence even on solely T1w studies, with some reporting increases in the volume of the globus pallidus [3, 61] and other work suggesting steady declines in globus pallidus volume throughout childhood/adolescence [2]. The different developmental trajectory observed for the globus pallidus measured on DTI compared to T1‐based segmentation in the current study may reflect the improved GM/WM contrast around the internal capsule provided by the FA + MD composite map. Specifically, the apparent volume reduction in T1w‐based segmentation during childhood likely reflects the ongoing myelination of the adjacent internal capsule reducing the GM/WM contrast, effectively making the boundary difficult to distinguish on T1w images. In contrast, the DTI‐based method remains sensitive to the underlying GM/WM boundaries, potentially revealing a volumetric growth phase during adolescence/early adulthood that is masked in standard structural imaging. This suggests the globus pallidus is a late‐maturing structure with constructive growth continuing later relative to the earlier developmental peaks reported in large‐scale T1w volumetric studies [3, 61]. The proposed DTI‐only segmentation provides an additional avenue for measuring deep GM volume; however, more work is needed to fully understand the true volumetric development and aging trajectories of the subcortical GM and the effects differing segmentation algorithms have on their measurement.

4.7. Sex Differences in Deep GM

Consistent with studies observing sexual dimorphism in total brain and subcortical volume [3, 62], males exhibited larger volumes across all five deep GM structures compared to females. However, the observation of higher MD in males across four of the five structures (without accompanying FA differences) suggests sex differences beyond dimorphic volumetric scaling. This may reflect subtle sexual dimorphism in microstructural cell packing density or fluid distribution within deep GM but requires further investigation. The similar lifespan trajectories observed between sexes suggest that although biological sex influences macro‐ and microstructural properties, it does not materially alter our understanding of the timing or rate of deep GM maturation and aging.

4.8. Study Limitations and Future Directions

Frequent errors were encountered when segmenting the amygdala using the proposed method. These errors were a result of (1) the anterior boundary separating the amygdala from the cortex was indiscernible on the FA + MD composite map and/or (2) large distortions caused by the adjacent air tissue interface corrupted the image in this entire region. Importantly, even with these errors, the volumes of the amygdala were within range of those reported by FreeSurfer, but even still, FA and MD results reported for the amygdala using the proposed method should be interpreted with caution.

An important consideration for future clinical application of the proposed method is the potential effect of pathological changes on segmentation accuracy. The surface deformation algorithm relies on image edge contrast rather than absolute FA or MD values to define structure boundaries; thus, the segmentation is expected to remain accurate provided sufficient contrast exists between deep GM and adjacent tissue. However, in cases of severe pathology, such as extensive neurodegeneration, neuroinflammation, edema, or WM lesions, the contrast at structure boundaries could be disrupted. For example, demyelination of the internal capsule could reduce the FA gradient between the capsule and adjacent basal ganglia, potentially leading to inclusion of WM voxels. Similarly, the FA and MD cutoffs applied during initialization (FA > 0.5, MD > 1.5 × 10⁻³ mm²/s), while lenient for healthy tissue, could theoretically exclude severely affected GM voxels with abnormally elevated MD. Furthermore, because the segmentation relies on the same DTI contrast from which diffusion parameters are subsequently extracted, any pathology‐driven segmentation error would systematically bias the measured FA and MD values; for instance, exclusion of damaged peripheral voxels would underestimate the extent of microstructural change. Future validation of this method in clinical populations should therefore include careful quality assessment of segmentations overlaid on DTI maps, ideally benchmarked against manual segmentation or alternative imaging contrasts, to quantify any pathology‐related bias in both the volumetric and diffusion measurements.

The imaging acquisition used in the current study was a notably higher resolution of 1.5‐mm isotropic relative to a typical 2‐mm isotropic whole‐brain diffusion acquisition. High resolution diffusion imaging studies in large cohorts are becoming more common; the diffusion protocol for the HCP development (HCP‐D) and aging (HCP‐A) cohorts is acquired at a 1.5‐mm isotropic resolution, and the proposed methodology ought to be applicable to these datasets as well [63].

Other work proposing whole‐brain segmentation on native diffusion images has demonstrated the advantage of using these segmentations to improve WM tractography [22, 24]; hence, the proposed method may also aid in this context. A unique aspect of the proposed method is that it generates 3D meshes of the subcortical GM structures that align with the GM/WM boundaries visible on the diffusion images/maps. Building upon this work, tractography algorithms have been able to utilize the shape of deep GM to segment the striatum's caudolenticular GM bridges [64] with future work planned to leverage the shape and size of deep GM structures to generate more anatomically plausible tractography reconstructions.

Modeling the diffusion signal in GM is still an open problem given the existence of molecular exchange between cellular and extracellular compartments [65]. In this study, we chose to represent the diffusion signal as a tensor to avoid assumptions about the underlying microstructure of the five deep GM structures of interest. In fact, other studies of development [7, 14] and aging [66] have proposed using higher order diffusion models to study the deep GM; however, to date, many of these higher‐order models are designed for WM and are unsuited to the microstructural environment of GM, so developmental/aging trajectories generated by these techniques should be interpreted with caution. Outside of the scope of this work, future studies will explore diffusion modeling techniques that are more tailored to the microstructural environment of the individual deep GM structures.

5. Conclusions

A fully automated segmentation method is proposed that is designed to work directly on diffusion images for the purpose of measuring DTI parameters from five deep GM structures (globus pallidus, striatum, thalamus, hippocampus, and amygdala). The segmentation uses the contrast available on DTI from the mean b1000 DWI and a FA + MD composite map to deform 3D meshes to the boundaries of each of the target structures. Compared to standard T1‐based FreeSurfer segmentation, the proposed method generated similar segmentations for the five deep GM structures and may outperform T1‐based segmentation methods due to the excellent GM/WM contrast visible on the FA + MD composite map throughout the internal/external capsule. In a typical development and aging cohort including 357 participants over the “lifespan” of 5–90 years, unique development and aging trajectories were observed for FA and MD of each deep GM structure that were distinct from their observed volumetric trajectories, suggesting that DTI may provide additional information about brain maturation in the deep GM. The automated DTI‐only segmentation does not require T1w imaging and has clear advantages over T1w‐based segmentation. Thus, the method will be useful for future studies of the diffusion properties of deep GM in development, aging, or clinical populations, where T1w images either add to the total acquisition time or are not acquired routinely.

Author Contributions

Graham Little: writing – original draft, conceptualization, methodology, data curation, formal analysis, software. J. Alejandro Acosta‐Franco: data curation, analysis. Christian Beaulieu: conceptualization, grant funding, methodology, writing – original draft, writing – review and editing.

Funding

This work was supported by the Canadian Institutes of Health Research, the University Hospital Foundation, Women's and Children's Health Research Institute, Canada Research Chairs, the Natural Sciences and Engineering Research Council of Canada, and the Unifying Neuroscience and Artificial Intelligence in Quebec.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Group average volume measurements for the HCP test–retest cohort extracted per structure using standard 3D T1‐weighted automated segmentation (FreeSurfer) and the proposed diffusion‐only segmentation method, along with Dice scores reflecting inter‐method voxel overlap.

Table S2: Reproducibility of proposed diffusion‐only segmentation method over two scans. Group average volumetric measurements for the HCP test–retest cohort extracted per structure using the proposed diffusion‐only segmentation method along with inter‐scan dice coefficients. No significant pairwise differences (p < 0.05) were observed between Scan 1 and Scan 2.

Figure S1: (A) Voxel‐wise segmentation generated from FreeSurfer (HCP standard structural processing pipeline) on the HCP T1‐weighted data (22‐year‐old male) overlaid on an axial slice from a T1w structural image. (B) Slice intersections of 3D models generated from the proposed DTI‐only segmentation method overlaid on the mean b1000 DWI from the approximately same slice relative to the T1w image in the same subject. (C) Slice intersections from the proposed DTI‐only segmentation method displayed along with FreeSurfer segmentations (nearest‐neighbor interpolation from anatomical image to diffusion data) overlaid on the FA + MD composite map. Segmentations are displayed for the globus pallidus (pink), striatum (green), thalamus (yellow), and hippocampus (cyan). In this example, the voxel‐wise FreeSurfer segmentation poorly delineates the edge of the four structures reflecting the poor T1w GM/WM contrast in regions adjacent to these structures.

Figure S2: Deep GM FA (Column 1), MD (Column 2), and volume (Column 3) calculated using the proposed DTI‐only segmentation method and plotted versus age separately for females (red) and males (blue) for all subjects from the University of Alberta Normative database (N = 156 males and 209 females). Trajectories followed the same general trends for males (blue line) and females (black line) for FA, MD, and volume. Four (globus pallidus, striatum, thalamus, and hippocampus) of the five deep GM structures had greater MD for males compared to females (albeit slight) and all five structures had a higher volume in males compared to females on average over the shown lifespan (* assessed by t‐test p < 0.001).

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.