CRL Diffusion MRI Atlas

• CRL Diffusion MRI Atlas is a comprehensive spatiotemporal fetal brain resource that integrates diffusion-weighted MRI with expert-refined labels for segmentation and morphometric analysis.
• It applies motion-robust slice-to-volume registration and log-Euclidean averaging to accurately derive fractional anisotropy, mean diffusivity, and principal diffusion directions despite fetal motion.
• The atlas uses diffeomorphic registration and kernel regression on age to enable population-level neuroanatomical studies and joint integration of multimodal data.

Searching arXiv for relevant papers on the CRL Diffusion MRI Atlas and closely related atlas-based diffusion MRI work. The CRL Diffusion MRI Atlas denotes a diffusion MRI atlas resource within the CRL atlas family, most explicitly as part of the CRL-2025 fetal brain release, where diffusion-weighted MRI and DTI atlases are distributed together with newly created tissue segmentations, labels, and segmentation tools for fetal brain MRI analysis (Bagheri et al., 20 Aug 2025). In the broader CRL-related atlas literature, diffusion MRI atlas construction is also tied to a multimodal framework in which structural connectomes are represented as Riemannian metrics and an atlas is defined as the Fréchet mean of a population, enabling joint integration of white matter structure from DWMRI and cortical structure from T1-weighted MRI (Campbell et al., 2021). Taken together, these works position the CRL diffusion MRI atlas as both a concrete reference resource and a methodological program for population-level diffusion-informed neuroanatomical analysis.

1. Scope and identity of the atlas

Within the CRL-2025 release, the atlas is a spatiotemporal fetal brain resource spanning 21–37 gestational weeks and constructed from carefully processed MRI scans of 160 fetuses with typically-developing brains, comprising 194 MRIs (Bagheri et al., 20 Aug 2025). The release includes T2-weighted anatomical atlases, diffusion-weighted MRI / DTI atlases, detailed multi-class tissue segmentations, 126-region anatomical parcellations, and atlas-to-subject automatic segmentation pipelines implemented through both multi-atlas and deep learning approaches. The same release states that the CRL-2025 atlas is accompanied by the CRL diffusion MRI atlas with newly created tissue segmentation and labels, making the diffusion component part of a larger segmentation and reference framework rather than an isolated DTI template.

A key point is that the diffusion atlas is not merely a scalar map repository. The release couples diffusion-derived templates with propagated and expert-refined labels, which makes the atlas usable for segmentation, morphometric analysis, and cross-modality studies. This suggests that the CRL diffusion MRI atlas is intended as an operational reference standard for fetal neuroimaging workflows, not only as a descriptive template.

2. Source data and diffusion-specific processing

The fetal diffusion atlas is derived from a cohort of 160 healthy singleton fetuses scanned between 21–37 weeks gestational age, primarily on Siemens 3T scanners (Bagheri et al., 20 Aug 2025). The diffusion acquisitions consist of 2–8 scans per subject at $b = 500 \, \mathrm{s/mm^2}$ with 12 directions. The associated anatomical component is based on T2-weighted HASTE MRI, which serves as the main dataset for the anatomical atlas.

Diffusion MRI preprocessing is described as motion-tracked slice-to-volume registration, motion-robust diffusion tensor estimation, two-stage registration comprising affine and nonlinear alignment, tensor reorientation, and log-Euclidean averaging. For the fetal setting, these steps are consequential because fetal imaging is particularly sensitive to motion and geometric inconsistency across slices. The atlas therefore reflects a pipeline in which reconstruction and tensor estimation are tightly coupled to motion compensation.

The diffusion measures provided by the atlas are fractional anisotropy, mean diffusivity, and principal diffusion directions. Tissue and regional segmentations are propagated from the T2-weighted atlas to DTI mean diffusivity maps through deformable registration and then refined by expert correction. A plausible implication is that the atlas is designed to support both microstructural interpretation through diffusion measures and anatomical interpretation through label transfer.

3. Spatiotemporal construction framework

The CRL-2025 atlas is explicitly described as spatiotemporal and 4D, with atlas construction based on a diffeomorphic deformable registration framework integrated with kernel regression on age (Bagheri et al., 20 Aug 2025). The registration model is specified as SVFFD, following Schuh et al., and the mean intensity or shape at each iteration is written as

$$\bar{I}_k(t) = \sum_{i=1}^{n} w_i(t) \left( \tilde{I}_i \circ T_{k,i}^{-1}(t) \right).$$

Here, the normalized subject image $\tilde{I}_i$ is mapped into atlas space by $T_{k,i}(t)$, and the temporal weights are normalized, kernel-truncated Gaussians in age. The mapping decomposition is given by

$T_{k,i}(t) = \bar{\phi}_k(t) \circ \phi_{k,i},$

with transformation averaging performed through a log-Euclidean mean. The release further states that recursive longitudinal deformation tracking allows generation of atlas instances at any age point in weeks or days.

For the diffusion atlas specifically, individual DTI volumes are built through motion-robust slice-to-volume reconstruction, then registered in two stages and averaged in the log-Euclidean domain. This makes the diffusion atlas temporally indexed rather than static. In practical terms, age acts as an explicit axis of atlas construction, which is essential for a fetal resource because tissue appearance and diffusion structure change rapidly across gestation.

4. Label system, anatomical coverage, and segmentation tools

The CRL-2025 release associates the diffusion atlas with a detailed labeling system comprising 36 tissue classes and 126 anatomical regions (Bagheri et al., 20 Aug 2025). The tissue labels include cortical plate; white matter with transient compartment segmentation; bilateral subcortical structures such as caudate, lentiform, thalamus, and hippocampus; cerebellum including newly added bilateral vermis; brainstem; and CSF spaces with new explicit labels for cavum septum and the third and fourth ventricles. The transient white matter compartments are subdivided into ventricular and intermediate zones and subplate up to 31 weeks gestational age, after which they are no longer MRI-distinguishable.

The 126-region parcellation is based on gyri, sulci, lobes, and other standard anatomical sub-regions, using the parcellation protocol from Blesa et al. 2016. Relative to CRL-2017, the new release reports sharper, higher-contrast anatomical templates, expanded anatomical detail, more accurate CSF and ventricle labels, bilateral vermis, and more rigorous manual validation and correction of automatic label propagation. It also reports improved anatomical sharpness both qualitatively and by median edge sharpness, with improvements from $+0.14$ to $+0.39$ across ages.

Two segmentation families accompany the atlas. The multi-atlas segmentation pipeline uses ANTS for rigid, affine, and symmetric-diffeomorphic registration, followed by label propagation and probabilistic STAPLE fusion; it is reported to be optimal when using age-matched atlases within $\pm 1$ week gestational age. The deep learning component evaluates nnU-Net, UNETR, SwinUNETR, and EMM-Seg. EMM-Seg is described as a U-Net backbone with Residual Vision MAMBA blocks and depthwise separable convolutions, with 2.89M parameters compared with 31M for nnU-Net, and training uses a hybrid loss combining voxel-wise cross-entropy and Dice loss. Validation on the BCH and FeTA datasets reports that EMM-Seg achieved the best results on major structures, in the Dice $0.85$–$0.77$ range, and outperformed in the presence of noisy or abnormal data.

5. Mathematical lineage in CRL multimodal atlas construction

A related CRL atlas-building line formulates the structural connectome as a Riemannian metric defined from diffusion tensors, in order to enable population-level statistical analysis of white matter architecture (Campbell et al., 2021). In that framework, the connectome metric may be written as

$$\bar{I}_k(t) = \sum_{i=1}^{n} w_i(t) \left( \tilde{I}_i \circ T_{k,i}^{-1}(t) \right).$$0

where $$\bar{I}_k(t) = \sum_{i=1}^{n} w_i(t) \left( \tilde{I}_i \circ T_{k,i}^{-1}(t) \right).$$1 is the diffusion tensor and $$\bar{I}_k(t) = \sum_{i=1}^{n} w_i(t) \left( \tilde{I}_i \circ T_{k,i}^{-1}(t) \right).$$2 is chosen so that the metric geodesics follow empirically estimated local fiber orientations. The space of Riemannian metrics $$\bar{I}_k(t) = \sum_{i=1}^{n} w_i(t) \left( \tilde{I}_i \circ T_{k,i}^{-1}(t) \right).$$3 is equipped with the Ebin metric

$$\bar{I}_k(t) = \sum_{i=1}^{n} w_i(t) \left( \tilde{I}_i \circ T_{k,i}^{-1}(t) \right).$$4

and the atlas is defined as the Fréchet mean

$$\bar{I}_k(t) = \sum_{i=1}^{n} w_i(t) \left( \tilde{I}_i \circ T_{k,i}^{-1}(t) \right).$$5

This construction is paired with diffeomorphic registration and can be extended to a joint multimodal energy that integrates both metric matching and T1-weighted image matching. The paper illustrates the framework with 2D examples and a 3D multimodal atlas built from T1 images and connectomes derived from diffusion tensors from six Human Connectome Project subjects. It reports that a metric-only atlas yields a blurry mean T1 image in gray matter, whereas adding T1 matching gives crisp gray matter anatomy without degrading, and possibly improving, the white matter metric atlas.

This line of work is relevant because it formalizes a longstanding controversy noted in the paper: whether tractography-derived structural connectomes can support quantitative population-level analysis. The proposed response is not to average tracts directly, but to move connectomic information into the space of Riemannian metrics, where geodesic distance, Fréchet means, and atlas construction become well defined.

6. Relation to other atlas-based diffusion MRI methods

The CRL diffusion MRI atlas should be distinguished from other diffusion-aware segmentation and atlas paradigms. A common misconception is to treat any MRI method that uses diffusion features as if it were built on the CRL atlas. For example, the thalamic nuclei segmentation method in “Domain-agnostic segmentation of thalamic nuclei from joint structural and diffusion MRI” explicitly does not use the CRL Diffusion MRI Atlas; instead, it relies on a public histological atlas together with a joint T1/dMRI Bayesian segmentation method to generate silver-standard labels for a CNN (Tregidgo et al., 2023). That method uses fractional anisotropy and the principal eigenvector encoded as an RGB image, accepts T1 and diffusion data of any resolution, and produces a 0.7 mm isotropic segmentation irrespective of input resolution, but its anatomical prior is histological rather than CRL-based.

Other atlas-based diffusion MRI methods illuminate the broader design space. Atlas-powered deep learning uses non-linear diffusion tensor registration to build biomarker atlases and atlas reliability maps, then combines the dMRI signal, a registered biomarker atlas, a registered atlas reliability map, and an atlas-to-subject alignment error map as inputs to a deep network for biomarker estimation from down-sampled diffusion MRI (Karimi et al., 2022). In adjacent medical image segmentation work, DiffAtlas replaces conventional image-to-mask mapping with a diffusion model over joint image-mask pairs and reports strong performance in limited-data and zero-shot settings on CT and MRI (Zhang et al., 9 Mar 2025). These methods are not CRL atlases, but they demonstrate a shared principle: atlas information can be made computationally active through registration, uncertainty modeling, or generative priors.

7. Uses and research significance

The fetal CRL diffusion MRI atlas is presented as a normative spatiotemporal reference for fetal brain structure and microstructure, with direct support for automatic segmentation and parcellation of both anatomical and diffusion scans through multi-atlas and deep learning methods (Bagheri et al., 20 Aug 2025). The release identifies regional morphometric, volumetric, and connectivity analyses; tract segmentation and tractography; investigation of developmental trajectories such as white matter maturation, regional brain growth, shape analysis, and FA or MD changes; and support for groupwise and population-level studies, including early biomarkers of atypical or pathological development.

The multimodal atlas framework based on Riemannian metrics extends this significance beyond fetal neuroimaging by enabling atlas-based tractography and population-level tract analysis in a setting where local orientations and long-range white matter pathways are preserved under geometric registration (Campbell et al., 2021). This suggests a broader interpretation of the CRL diffusion MRI atlas concept: a diffusion-informed reference space in which anatomical correspondence, microstructural descriptors, and population statistics can be studied jointly.

Because the resource is publicly released with open-source code, data access, labels, and segmentation models, it also functions as infrastructure. In that sense, its importance lies not only in the templates themselves, but in the combination of age-resolved atlas construction, expert-refined diffusion-space labels, and computational tools that make the atlas directly usable for segmentation, groupwise comparison, and developmental neuroimaging research.