A large-scale multi-species resource reveals cross-species generalization in label-free 3D neuron segmentation

This record contains a large annotated three-dimensional (3D) label-free neuronal soma dataset acquired from rat, mouse, and human brain tissue using oblique illumination and Dodt gradient contrast (DGC) microscopy.

The dataset comprises 240 volumetric image stacks across six species–modality combinations: rat oblique, rat DGC, mouse oblique, mouse DGC, human oblique, and human DGC. In total, the dataset includes 6,479 manually annotated 3D neuronal soma instances and 73,081 two-dimensional cross-sectional annotations. Each species–modality combination contains 40 volumetric stacks.

The dataset was created for systematic benchmarking of 3D neuron instance segmentation in label-free microscopy and supports evaluation of within-domain segmentation, cross-modality generalization, cross-species transfer, mixed-species training, and depth-dependent segmentation errors.

The uploaded archives contain the image volumes and corresponding consensus annotation masks for each species–modality combination. The annotations represent visible neuronal somata as observed in the respective label-free imaging modality, rather than complete neuronal morphology.

This dataset accompanies the manuscript:
"A large-scale multi-species resource reveals cross-species generalization in label-free 3D neuron segmentation."

Code for preprocessing, training, evaluation, and analysis is available at:
https://github.com/podtyazhki1337/3d-neurons-segmentation

Trained model weights are available separately on Hugging Face at:

https://huggingface.co/Podtyazhki1337/3d-neurons-segmentation

In file and folder names, "dodt" denotes Dodt gradient contrast (DGC) microscopy.