CUDA Deconvolution

We’ve recently been testing the graphics card accelerated deconvolution software from the Butte lab [1]. It’s very impressive – we can deconvolve a 1024 x 1024 x 50 slice image stack in about 8 seconds. ┬áThe test data we were using has some spherical aberration, so the resulting deconvolved images aren’t that nice and I won’t post them, but I think that’s the fault of our data and not of the software.

The data set size you can deconvolve is limited by the amount of memory on the graphics card, so the 1024 x 1024 x 50 data set fit fully into the graphics card RAM, a 1536 x 1024 x 50 data set required using some CPU RAM in order to deconvolve, and I was unable to process a 2048 x 2048 x 50 data set.

We’ve tried two different graphics cards; here is the time required to deconvolve the 1024 x 1024 x 50 data set if you are interested:

Quadro K2000: 12.5 sec
GTX 750Ti: 8.1 sec

I hope to do some more comprehensive testing and comparison of different deconvolution tools, but this one is the fastest of all the ones I’ve seen.

References

  1. M.A. Bruce, and M.J. Butte, "Real-time GPU-based 3D Deconvolution", Optics Express, vol. 21, pp. 4766, 2013. http://dx.doi.org/10.1364/OE.21.004766

4 thoughts on “CUDA Deconvolution

  1. Hi

    I’ve also tested against the Bruce/Butte algorithm using some data provided by George McNamarra. I tested using deconvolutionlab and also a custom algorithm (custom as in it combines some pre-existing approaches together). Some detailed commentary (especially on SA) in this blog entry.

    http://truenorth-ia.com/blog/?p=1

    Is your raw data open and available anywhere??

    Brian Northan

  2. Hi Brian –

    The non-circulant deconvolution looks pretty interesting. Are you looking for data sets to test deconvolution in general? I can definitely provide you with the data I was using for testing, but as I mentioned it’s not the best quality, optically. I’ve been hoping to retake some data with minimal aberration and compare multiple deconvolution algorithms (perhaps working with the Butte lab) but I haven’t had a chance to do so yet. If you want the data, shoot me an email (kurt.thorn@ucsf.edu) and I’ll send you a download link.

Leave a Reply

Your email address will not be published. Required fields are marked *