Cross-shaped spike localizations

[b-grimaud]

When looking into the coordinates of individual spikes for a given cluster, I noticed that when plotted the coordinates all fit into a cross pattern. That is, a lot of spikes have the same x value, oddly rounded with no trailing decimals, for different y values and vice versa.

Here it is displayed for an individual cluster :

With the matching table, notice the x values at exactly 2640.0 and the y values at exactly 960.0.

	times	cluster_id	x	y
14	885	0	2640.000000	957.088501
23	1362	0	2635.247314	960.000000
24	1433	0	2642.525146	960.000000
89	5678	0	2640.000000	966.294067
164	7457	0	2639.830811	960.000000
...	...	...	...	...
887389	11838406	0	2640.000000	952.806824
887501	11843033	0	2638.202393	960.000000
887570	11847925	0	2631.216797	960.000000
887571	11848357	0	2631.220703	960.000000
887636	11849502	0	2640.000000	955.113098

5411 rows × 4 columns

With PlotAll in Lightning :

With PlotNeighbourhood :

I checked in both the saved HDF5 and directly in the spikes DataFrame, so I don't think this is an issue with saving/storing data.

I took a look into SpikeLocalizer, I'm not too familiar with C but I couldn't see anything obvious.

[b-grimaud]

Small update : localizations are just fine using simulated data as specified in the demo notebook.

Here it is with SpikeInterface simulation, with the probe enlarged to 64 columns :

Here it is with data from MEArec from the ephys test data repository :

However, the issue shows up again when I run it on data simulated with SpikeInterface, but with the probe object loaded from an actual 3Brain recording :

So this might be 3Brain-specific.

[mhhennig]

What is the channel pitch on your system (60um?)? This happens when it is larger than the radius around which the COM algorithm looks to estimate location. You can increase neighbor_radius and inner_radius. inner_radius is the important parameter, this determines how far channels can be away from the peak channel. neighbor_radius matters for removing duplicates. Does this help?

[mhhennig]

Aside, on an array with 60um or larger pitch the spatial information isn't great - it is still enough to differentiate between neurons that sit on or between channels, but the spatial signal is quite noisy.

[b-grimaud]

Ran a few tests and it was indeed an issue of pitch. I looked back on older files and the localizations were OK using the legacy detection algorithm and the built-in reader.

Turns out the SpikeInterface extractor for BioCam data doesn't properly infer electrode pitch, unlike the old reader. Specifying proper pitch and size (in my case 42 and 21 µm respectively) solves the issue.

Thanks a lot !

[mhhennig]

Hmm, the extractor should work. Could you create an issue about this? What's your system? It works fine for us for a recent BioCam X.

[b-grimaud]

I'll look into a bit further, we're using a BioCam Duplex with 3rd generation Accura chips.

[b-grimaud]

Nevermind, looks like 3Brain updated their chip nomenclature. The chips we knew as Accura do have a pitch of 60 µm, and it is correctly listed in the probeinterface defaults. The Arena chips are the ones with a pitch of 42 µm. I'll try increasing inner_radius a bit !

[b-grimaud]

So I tried a few things, and increasing inner_radius did not change anything even when raised to abnormally high values (up to 500 µm). Interestingly, arbitrarily changing the pitch to anything below 50 µm works, but as soon as it is equal or above to 50 the crosses appear again.

[mhhennig]

Hmm - have you also increased neighbor_radius? If this does not help I'll need to take a look!

[b-grimaud]

I had tried it once but not as thoroughly. It seems to work properly when the values are both scaled appropriately, e.g. a 60 µm pitch with an inner_radius of 90 µm and a neighbor_radius of 120 µm. Thanks !