requirements training data and transferability of model

[lschm]

Hi,

just looking into denoising options for calcium data and am left with a couple of questions regarding deepcad:

1. As I understand you need high frame rates to estimate the noise parameters - I have some recordings > 40 Hz but mainly sample at roughly 15 Hz. Is 15 Hz too slow to train on? Should it be possible to train on the faster data and then use to denoise the slow data? I have been using GCaMP8m.

2. In earlier issues there were some comments about training new models for different data sets. I have a number of mice with a number of different field of views, but with the same structures labelled. Should a trained model just be transferrable between different animals? What/when is there a case for training a new model?

Best wishes & thanks!

[cabooster]

Hi! Sorry for replying late.

1. It's totally correct to train models on the faster data and then use them to denoise the slow data. According to our experience, 40Hz and 15Hz are enough for denoising GCaMP7. Since GCaMP8 is faster than GCaMP8, I think training models with 40-Hz data is better.
2. There's no need to train models for each mouse. I think you can use one model for all data since the same structures are labeled with the same indicator. However, I think a better solution is to select representative stacks from all mice and all field-of-views to form a training set, so that you can train a model with good generalization.