CRL-2025 Fetal Brain Atlas

• CRL-2025 Fetal Brain Atlas is a spatiotemporal MRI reference that maps in‑utero brain development from 21 to 37 GW using weekly templates, 36 tissue labels, and a 126-region parcellation.
• It employs advanced SVFFD registration with age-kernel regression to achieve improved edge sharpness and temporal consistency compared to earlier atlases.
• The atlas integrates a companion diffusion MRI dataset and provides segmentation tools, including MAS and deep learning models, to support robust multi-modal fetal brain analysis.

Searching arXiv for the cited CRL-2025 atlas and closely related fetal atlas/segmentation papers to ground the article. The CRL-2025 Fetal Brain Atlas is a spatiotemporal (4D) MRI atlas of typical in‑utero brain development from 21 to 37 gestational weeks (GW), assembled from carefully preprocessed fetal MRI acquired at Boston Children’s Hospital (2014–2023). It provides weekly age-specific templates, 36 tissue labels, delineation of transient white matter compartments up to 31 GW, and a 126-region parcellation, together with a companion diffusion MRI atlas labeled in the same anatomical framework. The release also includes multi-atlas segmentation (MAS) tools, deep learning–based multiclass segmentation models, and pipelines for single-subject and groupwise analyses, positioning the atlas as a reference framework for spatial normalization, automatic segmentation and parcellation, volumetrics, morphometrics, cortical parcellations, tract segmentation, tractography, and multimodal developmental analysis (Bagheri et al., 20 Aug 2025).

1. Definition, scope, and advances over earlier CRL resources

CRL-2025 was introduced to establish a high-quality, temporally consistent fetal brain atlas for analysis across mid to late gestation. Its stated purpose is to support spatial normalization, automatic segmentation/parcellation, and normative developmental analysis. In contrast to the earlier CRL-2017 atlas, the new release emphasizes higher anatomical detail, broader labeling, improved image quality, and modern reconstruction and registration pipelines (Bagheri et al., 20 Aug 2025).

Several concrete advances are specified. First, anatomical detail and sharpness are improved, with median edge sharpness improved by +0.14 to +0.39 across ages. The paper gives representative comparisons: at 21 GW, CRL-2017 = 1.008 versus CRL-2025 = 1.349; at 35 GW, CRL-2017 = 0.605 versus CRL-2025 = 0.799. Second, label granularity is expanded to include 36 tissue labels, transient white matter compartments (WMC) up to 31 GW, and 126 regional parcellation labels. The release also refines CSF and cerebellar labeling through the addition of the cavum septum, third/fourth ventricles, and bilateral vermis, while removing or integrating small low-contrast labels such as the hippocampal commissure and subthalamic nucleus for robustness (Bagheri et al., 20 Aug 2025).

A further distinguishing feature is the paired diffusion MRI atlas, which is released with a newly created tissue segmentation and labels in the same anatomical framework. This makes the atlas explicitly multimodal at the resource level, even though the primary structural atlas is T2-weighted. In that sense, CRL-2025 consolidates template construction, label definition, segmentation tools, and downstream analytic infrastructure into a single atlas ecosystem rather than only distributing average images (Bagheri et al., 20 Aug 2025).

This suggests that CRL-2025 is best understood not merely as a template series, but as a coordinated reference space for structural, diffusion, and segmentation-driven fetal brain analysis.

2. Cohort, gestational-age coverage, and imaging protocols

The atlas was constructed from 160 fetuses and 194 MRIs, with some longitudinal sampling: 129 single-visit, 28 two visits, and 3 three visits ≥6 weeks apart. Subjects were selected within ±1 week windows around target atlas ages. Inclusion criteria required pregnant participants aged 18–45, no MRI contraindications, and exclusion of multiple gestations or any suspected/known fetal brain abnormality (Bagheri et al., 20 Aug 2025).

The structural atlas spans 21–37 GW, and weekly spatiotemporal templates are provided. The underlying framework can generate atlases at continuous ages, but the distributed atlas grid is weekly. The companion diffusion atlas spans 22–38 GW. The paper notes that fewer subjects at the extreme ages may reduce kernel support, and that transient white matter compartment labels are available only up to 31 GW because these compartments become indistinct later in gestation (Bagheri et al., 20 Aug 2025).

Structural MRI acquisition relied on T2-weighted HASTE. Reported scanner distribution was Siemens 3T Trio (n=5), Skyra (n=134), Prisma (n=53), Vida (n=1), plus one Philips 1.5T scan. Coils were 18-channel body matrix until Aug 2017 and 30-channel thereafter. T2-weighted acquisition parameters were reported as TE = 115–120 ms, TR = 1400–1600 ms, flip = 120–160°, slice thickness = 2–3 mm, and in-plane resolution = 1 mm, with matrix sizes 256×256–320×320, field of view 256–320 mm², and 10–20 minutes acquisition per session (Bagheri et al., 20 Aug 2025).

The diffusion protocol used 2–8 stacks per subject along orthogonal planes, with 1–2 b=0 and 12 directions at b=500 s/mm² per stack, TR ≈ 3000–4000 ms, TE = 60 ms, in-plane resolution = 2 mm, slice thickness = 2–4 mm, and 10–15 minutes per session. Motion handling depended on multi-stack fast 2D acquisitions, followed by retrospective slice-to-volume reconstruction (SVR) and motion-corrected dMRI reconstruction (Bagheri et al., 20 Aug 2025).

A useful contextual comparison is that age-conditioned fetal atlas usage has also been exploited for segmentation in later work such as AtlasSeg, which pairs each subject with the nearest gestational-age template from the “Spatiotemporal Atlas of the Fetal Brain” and shows improved segmentation stability at early and late gestational ages (Xu et al., 2024). That comparison is relevant because it clarifies that CRL-2025 belongs to a broader class of GA-stratified reference spaces, while offering denser labels and a paired diffusion resource.

3. Construction framework and temporal modeling

CRL-2025 uses a temporally consistent 4D atlas built with stationary velocity free-form deformation (SVFFD) diffeomorphic registration (MIRTK) integrated with age-kernel regression following Schuh et al. The method employs iterative Log-Euclidean averaging of transformations to enforce temporal coherence (Bagheri et al., 20 Aug 2025).

The age-continuous atlas is written as

$$A(x,t) = \frac{\sum_i K_\sigma(t - t_i)\, I_i(\phi_i^{-1}(x))}{\sum_i K_\sigma(t - t_i)}.$$

The Gaussian kernel is

$$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$

with normalized weights

$w_i(t)=g_i(t)/\sum_j g_j(t),$

and

$g_i(t)=\frac{1}{\sigma_t\sqrt{2\pi}\exp\!\big(-(t_i-t)^2/(2\sigma_t^2)\big).$

The paper further states that $\sigma_t$ adapts to local age density, and that weights are truncated outside a kernel width of one week to enforce locality. This produces an atlas that is continuous in age while remaining locally anchored to nearby gestational windows (Bagheri et al., 20 Aug 2025).

The iterative update is specified through a mean intensity estimate

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

with globally normalized $\tilde{I}_i$ and

$$T_{k,i}(t)=\bar{\varphi}_k(t)\circ\varphi_{k,i}(t).$$

The subject-to-atlas warp at observed age $t_i$ is

$\phi_{k,i}=\exp(\mathbf{v}_{k,i}),$

where $$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$0 minimizes the SVFFD energy registering $$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$1 to $$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$2. Longitudinal deformation from $$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$3 to $$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$4 is

$$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$5

and age-adjusted subject transformations are

$$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$6

Residual atlas deformation is computed as the Log-Euclidean mean of SVFs,

$$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$7

The registration energy is given in illustrative LDDMM/SVFFD form as

$$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$8

where $$K_\sigma(\Delta t) = \exp\!\big(-\Delta t^2 / (2\sigma_t^2)\big),$$9 is the time-varying velocity field, $w_i(t)=g_i(t)/\sum_j g_j(t),$0 is a differential smoothing operator, and $w_i(t)=g_i(t)/\sum_j g_j(t),$1 is a data term such as normalized cross-correlation or SSD. Optimization proceeds via a multi-resolution schedule with gradient-based updates under MIRTK defaults, and transformations are composed through the Baker–Campbell–Hausdorff approximation to remain within SVFFD (Bagheri et al., 20 Aug 2025).

This construction strategy places CRL-2025 within the line of spatiotemporal fetal atlases that seek not only sharp weekly templates but also temporal coherence. Related work on continuous implicit atlasing has addressed temporal inconsistency from a different angle, for example by representing the atlas as a continuous neural field over space and age (Chen et al., 2022). CRL-2025, however, retains an explicitly diffeomorphic, kernel-regressed, population-atlas formulation grounded in classical atlas-building practice (Bagheri et al., 20 Aug 2025).

4. Label system, tissue compartments, and diffusion atlas

CRL-2025 provides three levels of anatomical representation. The first is tissue segmentation, reported as 36 labels. These include the cortical plate vs developing white matter boundary, deep nuclei such as lentiform and caudate, internal capsule, thalamus, hippocampus, brainstem, cerebellum including bilateral vermis, major CSF spaces including ventricles, 3rd, 4th, and cavum septum, and the corpus callosum (Bagheri et al., 20 Aug 2025).

The second level is the delineation of transient fetal white matter compartments up to 31 GW. These are explicitly listed as cortical plate, subplate, intermediate zone, and ventricular and subventricular zones, with the paper noting that these compartments become indistinct and disappear thereafter (Bagheri et al., 20 Aug 2025).

The third level is a 126-label regional parcellation, spanning bilateral cortical gyri/sulci/lobes and subcortical regions, following the Blesa et al. parcellation. The paper further notes that the label hierarchy enables cortical parcellations by crossing tissue segmentation with regional labels (Bagheri et al., 20 Aug 2025).

Label generation combined propagation and expert editing. The process began with ANTs symmetric diffeomorphic registration from CRL-2017 age-matched atlases, followed by MAS with Probabilistic STAPLE, then multi-round expert refinement in ITK-SNAP under neurologist and neuroanatomist supervision. Each atlas required approximately 1 hour automatic processing + 2–4 hours manual editing (Bagheri et al., 20 Aug 2025).

The companion diffusion MRI atlas was reconstructed using motion-tracking slice-to-volume registration, then aligned via affine + non-linear registration to age-specific templates. Tensors were reoriented with finite strain, and a log-Euclidean tensor average produced a spatiotemporal DTI atlas across 22–38 GW. Derived maps include color FA, FA, MD, RD, AD, and principal diffusion directions (Bagheri et al., 20 Aug 2025). The reported formulas are

$w_i(t)=g_i(t)/\sum_j g_j(t),$2

$w_i(t)=g_i(t)/\sum_j g_j(t),$3

$w_i(t)=g_i(t)/\sum_j g_j(t),$4

The FA formula is printed in the source with a truncated denominator, but FA is explicitly listed among the derived maps (Bagheri et al., 20 Aug 2025).

Label transfer to diffusion space used diffeomorphic registration of the CRL age-equivalent T2w atlas to the DTI atlas via mean diffusivity, followed by propagation of tissue classes and regional parcellations and then manual refinement by experts. The paper notes that tract delineations and tractography can be performed using these labels and DTI principal directions, and cites a related white matter tract atlas (60 tracts, 23–36 GW) as already available in related work (Bagheri et al., 20 Aug 2025).

This multimodal label strategy is significant because it places the structural and diffusion atlases into a shared anatomical framework. A plausible implication is that CRL-2025 can support analyses that move directly between volumetric tissue anatomy, regional parcellation, and tract-level diffusion organization without redefining the coordinate system or label ontology at each stage.

5. Segmentation tools and computational workflows

The release includes both multi-atlas and deep learning segmentation tools. The single-subject T2w pipeline is described as follows: acquire multi-stack HASTE, perform SVR with SVRTK (defaults) or NiftyMIC (alpha=0.04), apply N4 bias correction and intensity normalization, perform brain extraction (intracranial cavity segmentation), rigidly align to age-matched atlas space with FLIRT, then run either MAS or DL inference. In the MAS path, age-matched CRL-2025 atlases within ±1 week are registered to the subject with ANTs SyN (cross-correlation), labels are propagated, and fused with Probabilistic STAPLE. Optional manual refinement is performed in ITK-SNAP, and cortical parcellations can be derived by crossing tissue and regional labels (Bagheri et al., 20 Aug 2025).

For deep learning, the paper evaluates nnU-Net, UNETR, Swin-UNETR, and EMM-Seg. Input modality is reconstructed T2w volumes with training labels from MAS with manual refinement. Two datasets were used for model evaluation: BCH, with 177 subjects, 20–38 GW, and labels for 31 regions; and FeTA, with 80 subjects, 20–37 GW, and 7 tissue classes (GM, DGM, WM, BS, CB, CSF, VEN) (Bagheri et al., 20 Aug 2025).

The training setup is reported precisely: a hybrid loss of cross-entropy plus Dice,

$w_i(t)=g_i(t)/\sum_j g_j(t),$5

with AdamW, lr = 1e-4, PolyLR schedule, batch size = 1, 200 epochs, PyTorch, and NVIDIA RTX A5000 GPU (Bagheri et al., 20 Aug 2025).

Reported test-set performance is as follows:

Model	BCH DSC / HD	FeTA DSC / HD
nnU-Net	0.853±0.090 / 0.984±0.44	0.757±0.161 / 3.304±2.877
UNETR	0.821±0.101 / 1.33±0.64	0.764±0.165 / 3.941±2.934
Swin-UNETR	0.845±0.093 / 1.04±0.41	0.766±0.159 / 3.255±2.810
EMM-Seg	0.835±0.094 / 1.13±0.45	0.773±0.150 / 3.131±2.792

The paper states that performance is strong on BCH and somewhat reduced on FeTA because of lower-quality reconstructions and occasional severe abnormalities, and characterizes EMM-Seg as offering good robustness and efficiency (Bagheri et al., 20 Aug 2025).

CRL-2025 also sits within a wider shift toward atlas-informed and age-conditioned segmentation. AtlasSeg, for example, uses a dual-U-Net with a gestational-age-matched atlas image and atlas label as an explicit anatomical prior and reports Average Dice: 0.9172 ± 0.0297 for cortical plate segmentation in-site (Xu et al., 2024). Such work does not define CRL-2025 itself, but it clarifies how CRL-2025’s weekly templates and labels can function as inputs to atlas-guided learning systems rather than only as passive registration targets.

6. Quantitative evaluation, applications, and limitations

CRL-2025 is quantitatively evaluated at several levels. At the atlas-image level, median edge sharpness is higher than CRL-2017 at all ages, including examples such as 28 GW: 0.883 vs 1.160, 34 GW: 0.632 vs 0.848, and 37 GW: 0.622 vs 0.764. At the segmentation level, boxplots across 31 regions on BCH show high DSC for cortical plate, subplate, CSF, ventricles, corpus callosum, and thalamus, with lower performance on smaller or more challenging structures such as amygdala, caudate, and fornix (Bagheri et al., 20 Aug 2025).

The intended applications are broad. The atlas is explicitly positioned for segmentation/parcellation of individual fetal MRIs, atlas-based label propagation, groupwise analysis, volumetrics, morphometrics, cortical parcellations, tract segmentation/tractography, multimodal analyses, and normative microstructural trajectories. Example groupwise analysis is expressed through age-localized kernel regression of regional volumes,

$w_i(t)=g_i(t)/\sum_j g_j(t),$6

with uncertainty estimable by bootstrap resampling or local weighted variance under the same kernel (Bagheri et al., 20 Aug 2025).

The materials are publicly released through Harvard Dataverse, including the T2w atlas DOI, the DTI atlas DOI, and code repositories for both MAS and deep learning segmentation. Weekly atlases are distributed as NIfTI volumes and label maps, with label key text files for visualization in ITK-SNAP (Bagheri et al., 20 Aug 2025).

Several limitations are made explicit. Motion, artifacts, and variable stack quality remain limiting; reconstruction fidelity depends on the number and orientation of usable stacks. Age coverage is restricted to 21–37 GW for T2w and 22–38 GW for DTI, with fewer subjects at extreme ages. Deep models show reduced performance on low-quality or markedly abnormal anatomy, and MAS remains recommended when geometry deviates substantially from typical development. The atlas represents typically developing brains, so the paper advises caution in clinical decision-making for atypical cases (Bagheri et al., 20 Aug 2025).

These limitations are consistent with a broader literature in which age-conditioned atlases and conditional neural atlas models have been developed specifically to handle low-data or pathological settings, including ventriculomegaly and agenesis of the corpus callosum (Dannecker et al., 11 Jun 2025). CRL-2025 itself is normative rather than pathology-specific, but the release structure suggests clear interoperability with such conditional atlas frameworks.

7. Position within fetal brain atlas research

CRL-2025 extends a lineage of spatiotemporal fetal atlas work represented in the paper’s cited resources, including Gholipour et al. (2017) and Serag et al. (2012), while adding a denser label system, diffusion coupling, and released segmentation tools (Bagheri et al., 20 Aug 2025). Its methodological center remains a diffeomorphic, kernel-regressed, temporally coherent population atlas, not a purely discriminative segmentation benchmark and not a purely neural implicit representation.

That distinction matters because contemporary fetal atlas research spans several paradigms. AtlasSeg operationalizes GA-matched atlas guidance inside a segmentation network (Xu et al., 2024). CAS-Net jointly learns conditional atlas generation and segmentation with age-bin conditioning (Li et al., 2022). CINA and CINeMA pursue registration-light or registration-free implicit neural atlases that are continuous in age and can be conditioned on anatomy or pathology (Dannecker et al., 2024, Dannecker et al., 11 Jun 2025). CRL-2025 differs by delivering a curated, explicitly labeled, publicly released atlas-and-toolkit resource whose construction remains grounded in classical deformable atlas-building and expert label refinement rather than in end-to-end generative representation learning alone (Bagheri et al., 20 Aug 2025).

A common misconception is to treat CRL-2025 as only an updated image template. The released resource is broader: it includes weekly structural atlases, a diffusion atlas, 36 tissue labels, transient white matter compartments, a 126-region parcellation, MAS, deep learning segmentation models, and example workflows for single-subject and groupwise analysis (Bagheri et al., 20 Aug 2025). Another possible misconception is to assume that the deep models supplant atlas-based methods in all settings. The paper is explicit that MAS remains recommended when anatomy is severely abnormal or motion/artifact is extreme (Bagheri et al., 20 Aug 2025).

In practice, CRL-2025 functions as a normative reference system for fetal MRI analysis between 21 and 37 GW, with explicit support for segmentation, parcellation, morphometry, and diffusion-informed analysis. Its main contribution is not only improved anatomical sharpness relative to CRL-2017, but the consolidation of template construction, expert-vetted labeling, multimodal alignment, and publicly released analysis tools into a single spatiotemporal atlas framework (Bagheri et al., 20 Aug 2025).