correcting motion/drift in 3D

[chiyu1203]

Dear Spikeinterface community,
I use 128-channel 4-shank-3D probes to do acute (typically 1-2 hour) recording in head-fixed, walking insects. Each shank (60um width) is 150 um along anterior-posterior axis and/or 300 um medial-lateral axis away from other shanks, and embedded with 32 channels for 3 columns (the interval between columns is 18.5um) spanning around 315 um along the dorsal-ventral axis.

Following the Handle motion/drift guide, I managed to detect peaks and estimate motion/drift in different methods. Since the span of the channels are much shorter than Neuropixels, I first tested different length of win_step_um and of win_scale_um (from 25um to 150 um). Then I focus on the analysis with DREDge because (a) It seems that DREDge and its ancestor detect more peaks than others

(figure removed to avoid confusion)

(b) After viewing @cwindolf wonderful talk and the paper, I feel confident that this method would work well with short probes (imaging these 3D probes as a multi-shank Neuropixels UHD but with NP1-Like geometry)

(figure removed to avoid confusion)

However, after DREDge detected those peaks, it does not look like those drift were corrected. I wonder that is because (a) the drift is too small to correct or the correction is too small to view (b) the drift occurred at the 3rd dimension, which is not detected in the function?! If (b) is the case, is there anything I can help develop this new feature? or should I split the stacked probes into 2 portion and analyse them independently?

Below is the code I use to call the function. Feel free to let me know if you have any thought.

recording_corrected, motion, motion_info = spre.correct_motion(recording=recording_saved, preset=preset, folder=folder,overwrite=False, output_motion=True, output_motion_info=True, estimate_motion_kwargs={'win_step_um':win_um,'win_scale_um':win_um},**job_kwargs)

[samuelgarcia]

Hi.

For small probe maybe you should use rigid mode, so no windowing.

It seems that DREDge and its ancestor detect more peaks than others

I do not anderstand this. peaks detection is one of the steps of drift estimation, you can control the threhold for it.
We changed ver time this threhold. The best is to have a high threhold to keep only the big spikes because the dirft is better estimated with biger ones. But in the other hands you need enought spikes so you can shuld the appropriate threshold you need for your data.

However, after DREDge detected those peaks, it does not look like those drift were corrected.

The drift you have is quite small (<5um), sometimes it is better to not correct when you have such small dirft.
The right/left ratser looks the same because you need to zoom a lot on the y scale to see the "correction".

You can always split your recording by shank and make one motion per shank maybe you will have sliftly diffrent results.

[chiyu1203]

Hi!!! Thank you so much for the reply. Sorry, I only found out the "analyse motion shank-by-shank" idea in this issue #3458 after I post that question. And thank you for pointing out changing thresholds! I am doing the analysis again and will update what I get soon.

[chiyu1203]

Hi I would like to give you an update about analysing motion/drift shank by shank in a 3D probe.
I used the abovementioned 3D probes to do acute recording in a fixed insect. The insect was placed anterior part facing the electrode and its limbs were cut off to stablise the recording. The prepartion looks similar to this paper, except that I do not have a pump to flow in saline during the recording (which caused a potential issue that I will mention later).
Then I inserted the multi-shank probes into different brain region, and recorded the spiking activity. When analysing motion/drift, I tried out different values for win_step and win_scale pamaters. The vertical length these 3D probes cover is similiar to Neuropixels UHD (but with NP1-Like geometry), so I tried dredge with parameters recommended for Neuropixels UHD and ended up using win_step=75, win_scale=150 for the rest of the analysis.
Below is the motion estimation and correction from each of the shank
Shank 0:

Shank 2:

Shank 3:

The first thing you can notice is that detected peaks vary across shank, which can be an issue due to the nature of the brain region and/or the threshold settings. The drift happened between 1500 and 2000 secs is probably due to the prepartion drying out during the recording and the probe started to move along the tissue. Interestingly, each probe has different dynamic of drift, which revealed how difficult correcting 3D drift can be.
In the end, I decided to stop using these probes and switched to longer probe with fewer shanks.
Hope you had good experience with your recording!

[cwindolf]

Hi @chiyu1203 , that's really interesting, thank you for sharing! Shank 3 in particular is cool to look at -- it seems like different units are moving in different directions at t~1750, which is not something that can really be modeled by a motion field that any of these algorithms consider. Good luck with the longer probes, and feel free to reach out if you have further questions or issues.

[chiyu1203]

No problem! Thanks for documenting those parameters in the first place!
I think that t~1750 drift happened as tissue dried out, the shank shifted both horizontal and vertically. This is why I can imagine the difficulty about 3D motion estimation with these kind of multi-shank probes. On one hand, these probes are made flexible so that they do not break easily when the tissue move. On the other hand, this flexibility makes it difficult to estimate the drift/motion across shanks as the physical distance among shanks is never fixed.