D-LMBmapX

Mapping how whole-brain circuitry forms during development is essential for understanding disorders of neural wiring, but the developing brain is a moving target: its morphology changes rapidly, varies from region to region, and differs substantially between individuals during early stages. No prior method could profile connectivity at arbitrary developmental time points, and the rapidly changing shape of developing axons makes segmentation especially hard. D-LMBmapX, from Jing Ren's group at the MRC Laboratory of Molecular Biology and posted to bioRxiv in February 2025, is a deep learning pipeline built to close this gap.

D-LMBmapX extends the earlier D-LMBmap connectivity-mapping toolkit to the developmental setting. It constructs sample-inferred atlases from flanking anchor stages so that any intermediate time point can be registered accurately, and — critically — it includes a foundation model for generalised axon and soma segmentation that works across developmental stages without per-stage retraining. Together these components enable quantitative mesoscale developmental connectomics.

The pipeline is demonstrated through spatial-temporal profiling of catecholaminergic projections and also supports robust cross-modality and cross-dimensional registration, including alignment of single 2D slices to 3D references, generalizing to broader 5D analysis of biomedical spatial datasets.

#Key Features

• Generalised segmentation foundation model: A single model segments axons and somata across developmental stages without retraining for each time point, handling the rapidly changing morphology of developing neurons.
• Sample-inferred developmental atlases: Atlases are constructed from flanking anchor stages, enabling accurate registration at arbitrary intermediate developmental time points.
• Cross-modality and cross-dimensional registration: Supports robust alignment across imaging modalities and dimensions, including precise registration of single 2D slices to 3D reference volumes.
• Mesoscale developmental connectomics: Enables quantitative whole-brain projectome analysis, demonstrated on the spatial-temporal profiling of catecholaminergic projections.

#Technical Details

D-LMBmapX is an end-to-end pipeline organized around three deep-learning modules — whole-brain axon and soma segmentation, brain-region segmentation, and multi-scale whole-brain registration with heatmap generation for density visualization. The segmentation stage is trained as a foundation model so that it generalizes across postnatal developmental stages rather than requiring annotated data for each stage, and the registration stage uses atlases inferred from flanking anchor stages to place any time point into a common reference. Pretrained models and a GUI application are distributed through the D-LMBmap software releases, with test datasets provided separately. The preprint has not yet completed peer review and is posted under a no-reuse (all rights reserved) license, while the accompanying D-LMBmap software is made available for research use.

#Applications

The pipeline is aimed at developmental neuroscientists and connectomics groups working with light-sheet or other whole-brain volumetric imaging. By segmenting axons and somata across stages and registering samples into stage-appropriate atlases, it lets researchers quantify how projections emerge and reorganize over development, compare individuals, and build mesoscale wiring maps. Its cross-dimensional registration also makes it useful for aligning heterogeneous 2D and 3D spatial datasets in broader biomedical imaging workflows.

#Impact

Developmental connectomics has been held back by the lack of tools that cope with the morphological variability of the growing brain. By casting axon and soma segmentation as a generalised foundation model and pairing it with stage-inferred atlases, D-LMBmapX enables whole-brain profiling at arbitrary developmental time points that previous pipelines could not address. As a preprint with a restrictive license, its uptake will depend on the released software and on peer review, but it points toward reusable segmentation backbones for developmental and 5D spatial imaging.