Replies: 5 comments 11 replies
-
Yea seeding with something besides the greedy init will work better here. If you have a method to segment your cells you can use that as your spatial seeds. |
Beta Was this translation helpful? Give feedback.
-
@cawarwick do you have the video for the image in the first post? |
Beta Was this translation helpful? Give feedback.
-
Hi, we're making some progress, it's a hard problem especially given the synchronous activity of the cells. Do you have ground truth for any of these movies? If you have more and longer movies that'd be useful too. How long are these videos typically, in number of frames. |
Beta Was this translation helpful? Give feedback.
-
We're getting better seperability, it's still a while before we can have an alpha for users to try out but if you've got more movies, longer is better (how long do you typically acquire for, longer is better): |
Beta Was this translation helpful? Give feedback.
-
I recently started imaging motor neurons, which similarly have a wide range of cell body sizes and a relatively high level of synchronization. I also haven't been able to find parameters that correctly identify the spatial footprints, though I've had some moderate success with seeding with my own masks that I've drawn in ImageJ. I am definitely interested in this pipeline that you all are working on! And I'm happy to share some of my data if that would be helpful at all. |
Beta Was this translation helpful? Give feedback.
-
I have recently started imaging the Dorsal Root Ganglia in which the primary sensory afferents reside. My problem arises from the huge variation in cell body size of these cells in which the largest neurons are >30um in diameter and the smallest are ~10um in diameter. The below image is from a single plane of a 2p recording showing the large variation with a range of around 2-3 fold. I have been messing with gsig and merge_thr to try and find an optimal balance but it feels like a push and pull that I can't get right. e.g. The small cells get merged together with too large of a gsig and so I increase the merge_thr, but then it starts to break up the large cells into 2 pieces. I've done a fair bit of hyperparamter searching and haven't found one I particularly like.
Not sure if there is an optimal solution, but I am not well versed in exactly how CNMF works beyond knowing how critical gsig is to finding cells since it determines the gaussian kernel with which to search for cells.
Thanks!
Beta Was this translation helpful? Give feedback.
All reactions