2p voltage imaging

[HakimBelm]

Hi,
I was wondering if you ever tested SUPPORT on Two photon Voltage Imaging dataset.
I am asking because I tried to train support on such datasets, and the resulting model failed at denoising my 2p voltage imaging movies. Basically, the resulting denoised video shows no difference/no improvement compared to the raw data.
Weirdly, when I try the pretrained model (bs1), I see a significant improvement in term of SNR, at the cost of a bit aggressive "smoothing" of the data, definitely reducing the absolute amplitude of deltaF/F peaks related to activity.
Best regards,

[SteveJayH]

Hi,

It sounds like the model might be learning an identity mapping — the denoised output being almost identical to the raw data can happen when the bs_size setting isn’t optimal. This occasionally occurs with motion-corrected datasets or imaging systems with pixel bleeding.

Although the default bs_size=3 usually works well, you can try increasing it to 5; that should help the model move away from the identity solution. However, note that a larger value can also make the output smoother, so it’s important to find the smallest setting that still avoids identity behavior.

If possible, please share the training script you used, and send a small demo TIFF so I can take a closer look (post here, or, email it to jay0118@kaist.ac.kr).

Regarding the pretrained bs1 model: it avoids the identity issue because the original training data didn’t have those artifacts, which is why it generalizes better in your case.