CALHippo builds a three-class, CA1–CA4-wide cell annotation library from 1 µm/px BigBrain sections and trains a UNet density model to infer cell density across the CA complex

CALHippo provides the first expert-validated, cell-level annotated dataset spanning all CA1–CA4 subfields with explicit three-class labels (excitatory neurons, inhibitory interneurons, glial cells). Existing hippocampal reconstructions relied on low-resolution data without explicit cell-type-resolved annotations, limiting quantitative maps of excitatory neurons, inhibitory interneurons, and glial cells; this dataset fills that annotation gap at the cell level. Evidence comes from newly released 1 µm/px BigBrain sections of the right hippocampus, with expert-validated annotations explicitly covering all CA1–CA4 subfields and three classes.

High-resolution cell instances are produced by a human-in-the-loop pipeline combining foundation-model-based segmentation, iterative expert correction, and model ensembling. Compared with purely automatic segmentation, iterative expert correction and model ensembling embed expert judgment in the annotation loop, making the production of cell instances and three-class labels traceable. The method description explicitly lists foundation-model-based segmentation, iterative expert correction, and model ensembling, and states that instances are classified into three cell classes.

By projecting sparse high-resolution annotations into 20 µm/px low-resolution BigBrain space to create class-specific supervision maps, a UNet-based density estimation model is trained to enable slice-by-slice inference across the full CA complex. This step extends sparse high-resolution annotations to the full volume, so information originally confined to local regions can support class-resolved density inference across the whole CA complex. The text states that supervision maps come from projecting high-resolution annotations into 20 µm/px space, that the model is a UNet density estimator, and that outputs are slice-by-slice density maps.

The predicted density maps are sampled to generate a class-resolved mesoscale cellular point cloud, and code and dataset are publicly released to support reproducibility. Beyond annotated data, the work also provides a mesoscale point-cloud map and public code and dataset links, making the resource reusable and reproducible. The text gives two public links, one for the code repository and one for the dataset, and states that the point cloud is sampled from predicted density maps.

The resource targets researchers who need cell-type-resolved hippocampal maps, in the setting of the right hippocampal CA complex: high-resolution annotations cover CA1–CA4 subfields, and density inference proceeds slice by slice in 20 µm/px low-resolution BigBrain space. It enables follow-up work to conduct circuit modeling and quantitative analysis on class-resolved mesoscale point clouds and density maps, and to reproduce or extend the pipeline via the public code and dataset.

The text is abstract-level and does not report annotation scale, cell counts, quantitative accuracy of density estimation, or validation metrics, nor the number of expert-correction iterations or how annotator agreement was assessed; those details require the original paper and the public dataset. In addition, density inference depends on projecting high-resolution annotations into 20 µm/px space, and the text does not quantify how much that projection affects the final point cloud.