Mindscope and BCI do this differently right now. Mindscope independently segments session ROIs and builds a mapping table based on ROI similarity heuristics.
BCI concatenates a handful of sessions with the same field of view and segments them together then re-uses these ROIs in subsequent sessions, but quality of ROIs degrades after ~1-2 weeks of sessions. Marton recommends moving toward a mapping table solution but is concerned about SnR for a single session based on his setup.
Points of contact:
Marton for AIND (Johannes to schedule)
Sam Seid for Mindscope (Arielle to schedule)
Marton's approach
motion correction
1a) construct a template for each session and align session
1b) align template of session n+1 with session n to obtain shifts S(n,n+1)
1c) obtain total shifts S(1,n+1) = S(1,n) + S(n,n+1)
1d) apply shifts -S(1,n+1) to align session n+1 with first template
find ROIs in concatenated movie
Currently uses rigid registration, which doesn't deal with the widening of the FOV due to heating (?)
Currently uses the same neural footprint for all sessions, which doesn't deal with changing footprints and blood-vessels
Rotation is not a concern
Tilted planes are of no concern (thrown out if it happens)
-> Piecewise rigid registration as implemented in CaImAn and Suite2p (not explicitly for cross-session) probably does the job
Could calculate neural footprints for each session to deal with changing footprints and blood-vessels
Most of the ROIs (>90%) are detected when considering a single session instead of all of them
z-shift is an issue!
In the long term, we could try to estimate z-shift dependent neural footprints using the recorded z-stacks (But are most of the neurons active when the z-stacks are recorded?)
Mindscope and BCI do this differently right now. Mindscope independently segments session ROIs and builds a mapping table based on ROI similarity heuristics.
BCI concatenates a handful of sessions with the same field of view and segments them together then re-uses these ROIs in subsequent sessions, but quality of ROIs degrades after ~1-2 weeks of sessions. Marton recommends moving toward a mapping table solution but is concerned about SnR for a single session based on his setup.
Points of contact:
Marton's approach
Currently uses rigid registration, which doesn't deal with the widening of the FOV due to heating (?) Currently uses the same neural footprint for all sessions, which doesn't deal with changing footprints and blood-vessels
Rotation is not a concern Tilted planes are of no concern (thrown out if it happens) -> Piecewise rigid registration as implemented in CaImAn and Suite2p (not explicitly for cross-session) probably does the job
Could calculate neural footprints for each session to deal with changing footprints and blood-vessels Most of the ROIs (>90%) are detected when considering a single session instead of all of them
z-shift is an issue! In the long term, we could try to estimate z-shift dependent neural footprints using the recorded z-stacks (But are most of the neurons active when the z-stacks are recorded?)
Papers: CaliAli SCOUT CellReg
Sam's Approach
From Marina for cross-session algorithm: https://github.com/AllenInstitute/ophys_nway_matching