jacobpennington
commented
1 month ago @EmmaCondon From your GUI screenshot, it looks like there's an issue with the data. There is a lot of coordinated noise across channels (notice the horizontal bands that all look very similar). That is most likely why a lot of the units in your phy screenshots have waveforms that don't look like spikes. You would need to address that problem with the data first, otherwise the sorting results are not going to be interpretable.
EmmaCondon
Describe the issue:
Hello, I am encountering a number of issues with KS4 output that I have not seen before. I would like clarification before moving forward with spike sorting. If you prefer I make these separate issues please let me know.
Background
I am new to KS4 GUI and Neuropixels, previously I have run automatic spike sorting using KS2 & KS3 and manually curated with Phy2 on the following probe types: Cambridge Neurotech 16-channel single shank, NeuoNexus 32- and 64-channel single shank, and NeuroNexus 128-channel multi shank probes. As I am new to KS4 and Neuropixels, I am unsure if what I am seeing is normal, but gauging from the Neuropixels slack and other issues here on github, my results do not look right. Additonally, these results are consistent across a number of recordings.
I am running KS4 via the GUI on acute recordings from NP1 probes (lasting approx 2 hours), not changing any parameters, and using phase3B1 channel map. Here is a view of the GUI after I load the recording. I have not encountered any errors while running KS4. Following automatic spike sorting using KS4 I am manually curating using Phy2.
Issue 1: Overmerging of units
During manual curation in Phy2 I am seeing very clearly in amplitude view and feature view (and sometimes in waveform view) that a unit labelled as good or MUA by KS4 is made up of two, separate units. You can see that after splitting, the autocorrelogram and waveforms of the units are different.
Here is another example before and after splitting.
However, in other cases when I split, the CCGs and ACGs are the same, indicating that they are the same unit.
Question: Could you please confirm that this is indeed overmerging and that there splitting here is the best course of action? I have alot of units like this and it is taking much time to split. My PI suggests that this is not actually 2 different units and splitting should not be done so clarification would be helpful.
Issue 2: Amplitude and PCs appearing as single line and dot
Another very common issue I am having is the appearance of the amplitude as a singular flat line, with a singular dot appearing in feature view. Here is an example,
Here are some examples:
And sometimes these line overlap, and when split are two different units but still appear as a line in amplitude and feature views.
Question: Are these units still considered acceptable? My thinking is that this is just a scaling issue with Phy but I have not seen this issue mentioned here in issues.
Issue 3: Waveform appearing on incorrect channel
In a number of different recording sessions, the channel number assigned to the unit does not match up in waveform view.
In the example below, you can see that this unit is labelled as 137 but in the waveform view, there is no waveform on 137, but it appears clearly on 133.
Here is another example, with channel 133 labelled in waveform view as 129.
Question: Is this due to an issue with the channel mapping? I hesitate to classify these labels as I do not want a mismatch in the channel numbers when analsying the data post spike sorting. I have seen an issue raised here (https://github.com/cortex-lab/phy/issues/866#issue-474267297) but this is related to KS2 and I am unsure if it can be applied to KS4 run via the GUI.
Issue 4: Units labelled as good by KS4 actually MUA
I know that it is common that incorrect labelling can occur in the automatic spike sorting process. However, well over half the units KS4 has assigned to as "good", as actually MUAs and have a clearly contaminated refractory period. For example, in a recording session in which KS4 has assigned 517 units as 'good', following manual curation Phy only 167 units are 'good' with the remainder having a refractory period contaminated by more than 20%.
Here are a few examples.
Question: Has this issue been reported elsewhere? It seems KS4 is over classifying units as 'good' even with clearly contaminated refractory periods. I have seen this previously while spike sorting (not using NP and using KS2) but never to the magnitude I am seeing now.
Reproduce the bug:
No response
Error message:
No response
Version information:
Operating System: Linux Ubuntu 18.04.6 Kilosort: 4.0.6 Python: 3.9.19 Phy: 2.0b5