CALHippo - Mapping neurons and glial cells in the human brain hippocampus in 3D using SOTA segmentation and density estimation models [R]
Hello everyone!
I'm posting our research work as you might be interested in how we used ML to map part of the brain cells of the human hippocampus :)
We used various human brain slices at high resolution (1 micrometer per pixel) and developed a custom segmentation pipeline that uses SoTA whole slice cell segmentation networks, like CellPoseSAM with good zero shot performances. We then refined semi-automatically those annotations and ensembled more finetuned models within the pipeline, adding a merging algorithm and a cell classification for 3 classes (excitatory and inhibitory neurons, and glial cells).
But the high-res slices covered only a few parts of the hippocampus with respect to other slices scanned at 20x less the resolution where the cell nuclei are only 1 pixel wide. So we tried to map the high-res annotations we obtained to the low-res corresponding slices, and used a small UNet to supervise a density estimation task for 3 classes. We obtained a network that outputs a density map that can be sampled to obtain a probabilistic map of the cellular positions.
Finally, to reconstruct the volume, we stacked together all the low-resolution density maps from all the slices that covered the hippocampus and obtained a point cloud, which you can see in the GIF along the corresponding anatomical CA (Cornus Ammonis) areas.
The performances are still limited by the quantity of data and low-resolution slices, but we showed that the results were biologically plausible given previous estimates by other researchers.
The paper was accepted at MICCAI 2026 a few weeks ago!
Feedback is very welcome, especially on the density-estimation formulation and possible uses of the generated point cloud.