Source-linked AI summary
A probabilistic atlas of the human thalamic nuclei combining ex vivo MRI and histology
Juan Eugenio Iglesias, Ricardo Insausti, Garikoitz Lerma-Usabiaga, Martina Bocchetta, Koen Van Leemput, Douglas N Greve, Andre van der Kouwe, Bruce Fischl, Cesar Caballero-Gaudes, Pedro M Paz-Alonso
TL;DR
Fine-grained thalamic nucleus segmentation is difficult because in vivo MRI may lack sufficient resolution and contrast, while registration-based histology atlases have limited population and contrast coverage. The paper builds a probabilistic atlas from reconstructed histology and in vivo anatomy, then applies it through Bayesian inference. The resulting tool showed agreement with prior histological volumes, excellent repeatability, robustness across MRI contrasts, and improved Alzheimer’s disease classification using subregional volumes.
Problem
Fine-grained thalamic nucleus segmentation is difficult when in vivo MRI lacks sufficient resolution or contrast, and existing registration-based atlases have limited variability and contrast coverage.
Method
The study built a probabilistic atlas from 26 nuclei in 12 histological thalami plus surrounding-anatomy segmentations from 39 in vivo scans, using MRI and blockface photographs for reconstruction.
Results
The atlas agreed with prior histological volume estimates, showed ICC scores mostly above 0.90 for test-retest reliability, and increased Alzheimer’s disease classification AUC by 0.20 using subregional rather than whole-thalamus volumes.
Takeaways & Limitations
The probabilistic atlas enables Bayesian segmentation of in vivo MRI scans across arbitrary contrasts and detects differential thalamic effects in Alzheimer’s disease.
Takeaways & Limitations
Alternative-contrast evaluation was positively biased because the same T1-derived aseg.mgz initialization was used, and some Alzheimer’s disease findings may partly reflect neighboring ventricular enlargement.
Abstract
from arXiv · showhide
The human thalamus is a brain structure that comprises numerous, highly specific nuclei. Since these nuclei are known to have different functions and to be connected to different areas of the cerebral cortex, it is of great interest for the neuroimaging community to study their volume, shape and connectivity in vivo with MRI. In this study, we present a probabilistic atlas of the thalamic nuclei built using ex vivo brain MRI scans and histological data, as well as the application of the atlas to in vivo MRI segmentation. The atlas was built using manual delineation of 26 thalamic nuclei on the serial histology of 12 whole thalami from six autopsy samples, combined with manual segmentations of the whole thalamus and surrounding structures (caudate, putamen, hippocampus, etc.) made on in vivo brain MR data from 39 subjects. The 3D structure of the histological data and corresponding manual segmentations was recovered using the ex vivo MRI as reference frame, and stacks of blockface photographs acquired during the sectioning as intermediate target. The atlas, which was encoded as an adaptive tetrahedral mesh, shows a good agreement with with previous histological studies of the thalamus in terms of volumes of representative nuclei. When applied to segmentation of in vivo scans using Bayesian inference, the atlas shows excellent test-retest reliability, robustness to changes in input MRI contrast, and ability to detect differential thalamic effects in subjects with Alzheimer's disease. The probabilistic atlas and companion segmentation tool are publicly available as part of the neuroimaging package FreeSurfer.
Introduction
The thalamus contains functionally distinct nuclei whose fine-grained in vivo segmentation is valuable but difficult because MRI resolution and contrast can be insufficient. This study addresses that gap with a probabilistic, histology-informed atlas designed for Bayesian segmentation across MRI contrasts.
- The thalamus comprises multiple nuclear masses with distinct functions, and finer histological subdivisions vary with classification detail.
- In vivo nucleus segmentation could support more specific morphometry, connectivity studies, surgical planning, and deep brain stimulation placement.
- Diffusion-based methods use local tissue properties or cortical connectivity to parcellate nuclei, but local diffusion information is typically insufficient alone.
- Registration-based histology atlases are limited because one subject-segmentation pair cannot represent population variability or register accurately across MRI contrasts.
- The proposed atlas combines histological nucleus delineations from 12 whole thalami with in vivo segmentations of surrounding anatomy from 39 subjects.
- Its probabilistic representation models surrounding anatomy and supports direct Bayesian segmentation of MRI scans with arbitrary contrast.
Materials and methods
The study used postmortem brains for ex vivo MRI and subsequent histological preparation, with MRI acquisition designed to preserve resolution and signal quality. Multi-slab artifacts were corrected computationally before histological reconstruction.
- Six postmortem cases supplied the ex vivo brains used to build the atlas.
- Ex vivo MRI was acquired at 3 T with a 12-channel coil while specimens were immersed in Fluorinert to reduce air-bubble and susceptibility artifacts.
- The multi-slab protocol preserved high-resolution acquisition but introduced slab-boundary artifacts at slab interfaces.
- Bias-field and slab-boundary artifacts were corrected simultaneously using a Bayesian method after brain-mask computation.
Histological analysis
Brains were sectioned into parallel blocks, photographed during microtomy, and sampled for Nissl-stained histology. Registration and classification procedures converted these materials into geometrically consistent tissue and histology data.
- Each hemisphere was cut into parallel 10–14 mm blocks spanning the frontal to occipital pole.
- One section every 0.5 mm was selected for Nissl staining and cytoarchitectonic analysis, while the remaining sections were preserved.
- Blockface photographs were registered and perspective-corrected to produce a geometrically consistent stack with known resolution.
- A random forest pixel classifier separated tissue from the block holder using visual features and training photographs from all cases and sides.
3D reconstruction of histology via blockface photographs
The reconstruction used blockface photographs as an intermediate target between histology and whole-brain MRI. Rigid block alignment and landmark-assisted nonrigid registration reduced distortions from sectioning, staining, and mounting.
- 3D reconstruction of histology via blockface photographs: Histology reconstruction required within-block nonrigid registration and between-block rigid registration.
- 3D reconstruction of histology via blockface photographs: Neighbor-only histology registration can produce geometric distortions such as the banana effect and accumulated z-shift.
- 3D reconstruction of histology via blockface photographs: Blockface photographs served as intermediate images because their section correspondence is known and they can be registered linearly to MRI.
- 3D reconstruction of histology via blockface photographs: Photograph stacks were rigidly aligned to whole-brain MRI using masked mutual information and iterative global and individual block transformations.
- 3D reconstruction of histology via blockface photographs: Section-level registration combined MRI intensity mutual information with B-spline transforms and manually placed corresponding landmarks to improve robustness to tears and folds.
Manual segmentation of nuclei on histology
An expert neuroanatomist manually delineated thalamic nuclei on digitized histology, then warped and refined these labels in MRI space to produce usable 3D segmentations.
- Manual delineation: An expert neuroanatomist manually delineated nuclei along the rostrocaudal thalamic axis using digitized histology and microscopy.The protocol followed Jones’s characterization of the human and mammalian thalamus.
- 3D reconstruction: Rigid and nonrigid deformations transferred the manual histological segmentations into MRI space, but gaps between blocks and adjacent-section inconsistencies required refinement.Blockface photographs and MRI served as registration references during reconstruction.
- Segmentation refinement: The refinement assigned labels to eroded and gap voxels by minimizing a cost function combining MRI intensity, distance from initial labels, and neighborhood smoothness.The method preserved labels inside the initial segmentation while encouraging locally coherent assignments.
- 3D reconstruction: The reconstruction workflow included reconstructed histology stacks and propagated segmentations before final label estimation.Examples show corresponding axial and sagittal MRI and histology views, while the refinement figure contrasts propagated and estimated labels.
Atlas construction
The atlas combined histological delineations of thalamic nuclei with in vivo segmentations of the whole thalamus and surrounding anatomy, using a Bayesian atlas-construction framework.
- Bayesian atlas construction: The atlas-construction method used Bayesian inference to estimate the probabilistic atlas most likely to have generated manual segmentations from in vivo and ex vivo data.This combines information from datasets acquired in different imaging contexts.
- Data integration: The atlas construction combined whole-structure segmentations from in vivo scans with reconstructed histological nucleus segmentations from ex vivo specimens.The surrounding structures included anatomy such as the caudate, putamen, and hippocampus.
- In vivo training data: The in vivo training data comprised 39 standard-resolution T1 scans with manual delineations of 36 structures, including both whole thalami.Scans were acquired at 1 × 1 × 1.25 mm resolution on a Siemens 1.5 T platform using MP-RAGE.
Segmentation of in vivo MRI
In vivo MRI segmentation was formulated as Bayesian inference using a warped probabilistic atlas and MRI intensity model, with tissue grouping chosen to improve robustness.
- Bayesian segmentation: The segmentation model warps the probabilistic atlas and uses a generative model of MRI scans within a Bayesian inference framework.The forward model provides the basis for estimating anatomical labels from MRI data.
- Parameter estimation: Model parameters were estimated by coordinate ascent, alternating deformation updates with Gaussian-parameter estimation.Deformations used conjugate gradients, while Gaussian parameters were estimated with expectation maximization.
- Tissue modeling: Grouping labels with similar intensity characteristics into shared tissue types improves algorithm robustness.The grouping accounts for similar MRI appearances among anatomical structures.
- Tissue modeling: Thalamic nuclei were grouped into three tissue-type sets, including reticular nucleus with white matter and mediodorsal and pulvinar nuclei together.The reticular grouping reflects its many crossing fibers and white-matter-like appearance.
- Evaluation: The atlas application was evaluated through four experiments addressing volumetric agreement and indirect segmentation performance.The evaluation included comparison with Krauth’s atlas and additional performance experiments.
Volumetric comparison with Krauth’s atlas
The proposed atlas produced representative-nucleus volume distributions that largely agreed with Krauth’s atlas, while direct whole-thalamus comparisons were limited by differing anatomical inclusion criteria.
- Compared nuclei: Six representative nuclei were compared across the proposed and Krauth atlases, including AV, LP, CM, MD, VL, and PU.MD, VL, and PU were defined as unions of their respective subdivisions.
- Comparison design: Volume distributions were compared across automated segmentations of 66 healthy adults rather than directly comparing the atlas templates.This avoided treating Krauth’s non-probabilistic mean as directly equivalent to the proposed probabilistic atlas.
- Volume results: Agreement between representative-nucleus volumes was high, although Krauth’s atlas yielded slightly larger bilateral AV and LP volumes and a slightly larger left PU volume.The distributions were visualized with violin plots.
- Scope boundary: Whole-thalamic volume comparison was not straightforward because Krauth’s atlas includes the red nucleus, subthalamic nucleus, and mammillothalamic tract, unlike the proposed atlas.The differing inclusion criteria also appear in the example segmentation comparison.
- Reliability: Whole-thalamus test-retest reliability exceeded 0.97 ICC, while representative-nucleus ICCs were all above 0.85 and most exceeded 0.90 or 0.95.These values were obtained from repeated scans acquired seven to 10 days apart in 31 subjects.
Robustness against changes in MRI contrast
The atlas-based segmentation remained broadly stable across MRI contrasts, with high whole-thalamus agreement and moderately high agreement for individual nuclei. T2 produced the least consistent segmentations, while synthetic MPRAGE scans with inversion times above 600 ms agreed best.
- Alternative-contrast segmentation used the same algorithm as T1 and initialized with FreeSurfer’s aseg.mgz segmentation.This shared T1-based initialization positively biases the robustness results, while matching the intended public-release scenario.
- Whole-thalamus segmentations achieved Dice overlap near or above 0.90 in almost all contrast comparisons.
- Individual-nucleus overlaps were approximately 0.75–0.85 for CM and VL, and approximately 0.65–0.75 for AV, MD and PU.
- Synthetic MPRAGE scans with inversion times over 600 ms showed the best cross-contrast agreement.Inversion times below 600 ms produce a contrast flip.
- T2 was the least consistent contrast because the boundary between MD/PU and the remaining nuclei was almost invisible.
Alzheimer’s disease study
The study evaluated whether atlas-derived thalamic measurements could distinguish Alzheimer’s disease from controls. Using all nuclei jointly substantially outperformed whole-thalamus volume alone, although the atlas-based whole-thalamus improvement was not statistically significant.
- The analysis compared recon-all whole-thalamus volume, atlas-summed whole-thalamus volume, and a leave-one-out LDA likelihood ratio.LDA incorporated all nuclei simultaneously while limiting dependence on stochastic classifier variation.
- Leave-one-out LDA using all nuclei achieved AUC = 0.830, versus AUC = 0.632 for FreeSurfer recon-all whole-thalamus volume.The comparison against recon-all was statistically significant at p ∼10^-10.
- Atlas-derived whole-thalamus volume achieved AUC = 0.645, slightly above recon-all’s AUC = 0.632, but the difference was not significant (p = 0.4).
- Fitting internal thalamic boundaries enabled some nuclei to separate Alzheimer’s disease and control subjects more accurately than whole-thalamus volume.The paper attributes this pattern to distinct regional connections and functions and unequal disease effects across nuclei.
- All nuclei showing significant group differences were larger in controls, using a Bonferroni-corrected threshold of p < 0.0019.
Discussion
The study introduced and evaluated a probabilistic atlas of 26 human thalamic nuclei, demonstrating reproducible segmentation across subjects and MRI contrasts and sensitivity to Alzheimer’s disease effects.
- Atlas and validation: The atlas combines 3D reconstructed histology from 12 thalami with Bayesian segmentation of in vivo MRI and surrounding anatomy.It was validated through four experiment sets.
- Atlas and validation: The atlas produced similar nuclear volume distributions to a previous histology-based atlas in 66 subjects.Its probabilistic formulation additionally supports scans with arbitrary MRI contrast.
- Reliability and contrast robustness: Test-retest reliability was excellent, with ICC scores mostly above 0.90 for scans acquired approximately one week apart.The result indicates that measured volumes were not primarily attributable to random fluctuations.
- Reliability and contrast robustness: Agreement between segmentations remained good across a wide array of MRI contrasts, supporting use of sequences with strong thalamic contrast such as FGATIR.The experiment assessed robustness to changes in input MRI contrast.
- Alzheimer’s disease findings: Alzheimer’s disease classification improved by 0.20 in AUC when using subregional thalamic volumes instead of whole-thalamus volume.Large effects were reported in mediodorsal, anteroventral, and ventral anterior areas.
- Alzheimer’s disease findings: The Alzheimer’s disease findings were consistent with neuroimaging and neuropathological reports implicating anterior and mediodorsal thalamic regions.The anterodorsal nucleus was included within the atlas’s anteroventral nucleus because of its small size.
- Interpretive caveats: The Alzheimer’s disease effects may partly reflect neighboring ventricular expansion, while reported lateral and medial geniculate atrophy could represent false positives or true disease-related effects.The authors identify small nuclei size and weak contrast with neighboring white matter as possible sources of false positives.
Conclusion
The authors provide a publicly available probabilistic thalamic atlas and segmentation tool for arbitrary-contrast in vivo MRI. They identify diffusion MRI integration as a key direction for improving segmentation and future disorder studies.
- Contribution and availability: The atlas is based on ex vivo MRI and histology, while the companion tool segments thalamic nuclei from arbitrary-contrast in vivo MRI.Both are presented as the paper’s principal outputs.
- Contribution and availability: Structural MRI-only segmentation supports analysis of large legacy datasets that lack diffusion data.The current Bayesian model relies on faint boundaries and prior knowledge to fit the atlas.
- Future improvements: Integrating diffusion MRI is proposed because local diffusion information and structural connectivity are strong signatures of divisions between thalamic nuclei.The authors expect these data to inform the generative model.
- Future improvements: Future analyses include relating nuclei to clinical scores and disease characteristics and investigating functional and structural thalamic networks.The proposed applications include Alzheimer’s disease, Parkinson’s disease, dyslexia, and schizophrenia.
- Contribution and availability: The tool is publicly available in FreeSurfer and is intended to support studies at sites without expertise or staff resources for manual 3D delineation.This extends access to automated thalamic nuclei analysis.