DEschweiler
commented
6 months ago Hi, thanks for reaching out to us. Unfortunately, this parameter is not yet implemented for the 3D variant. The main parameter that could be helpful in optimizing the results in that regard could be the "min_diameter" parameter, as it is also used to cluster all points after iteratively following the predicted gradient field. If that parameter is too small/large it can influence the splitting/merging of neighboring instances. We hope tweaking this parameter provides an easy solution, otherwise the combination of "min_diameter", "niter" and "step_size" all influence the clustering behavior. It could be that there is a different sweet spot of this triplet for the DAPI dataset than we determined for our datasets. If none of this helps to optimize the results, this is unfortunately indeed missing functionality of the current version.
Hi, I tried using the original Cellpose for 3D stack tif segmentation. The majority of the DAPI nuclei could not be identified in the first try. However, after adjusting the stitch_threshold from 0 to 0.2(I guess that >0 turn off 3D mode and use 2D mode for each slice, then stitch them together based on percentage of overlap), almost all the cells could be identified. Furthermore, I found that the larger this threshold is, the more likely it is for the DAPI ROI on neighboring slices from the same cell to be categorized as two different cells (based on the mask color or id number for that cell).
I am wondering if Cellpose3D also has the same threshold for adjusting the affiliation of two ROIs on neighboring slices to be the same cell or difference cells. Thanks.