Cell segmentation in 3D GCaMP structural recordings.
We segment cells using a 3-d fully convolutional network. Labels are generated using the red channel of GCaMP6 + mCherry expressing mice ('watermelon' mice). The network outputs two voxel-wise probability maps: one represents the estimated centroids of each cell and the other segments the cells; both are binary segmentations. To produce instance segmentations, we threshold the segmentation map and apply compact watershed using peaks in the centroid map like markers. See the demo for results. See the references for a similar approach applied to embryonic cells.
[1] Convolutional Neural Network-Based Instance Segmentation Algorithm to Acquire Quantitative Criteria of Early Mouse Development Yuta Tokuoka, Takahiro G Yamada, Noriko Hiroi, Tetsuya J Kobayashi, Kazuo Yamagata, Akira Funahashi bioRxiv 324186; doi: https://doi.org/10.1101/324186