accelerating the deconvolution step for long time series by re-using Fourier kernel?

[nikolas-claussen]

Hi there! First of all, thanks for developing this very helpful software.

I was wondering whether there might be a way of accelerating the deconvolution step for long time series.

Our lab recently began recording long & high time-resolution movies on our MuViSPIM. For these datasets, which can comprise up to 1000 timepoints, deconvolution becomes fairly time consuming, even on a GPU cluster. If I understand correctly, deconvolution works by iteratively applying transformations in Fourier space to compensate for image distortions, constructing the kernel by using the bead PSF as reference.

Would it be possible to save the Fourier kernel obtained for time point t, and re-use it as a starting point for time point t+1? Maybe one could get away with using a significantly smaller number of deconvolution iterations after the very first time point. Our movies have a time resolution of ~15s, so it seems likely that the same transformation should work for nearby time points, at least as the first iteration step. As far as I understand, the plugin currently does no such thing (all time points appear to take equally long to deconvolve).

How would I have to go about adding this feature, or is it likely a bad idea?