GPU Deconvolution WOW result on 2048x2048x32 plane Z-series ... formerly bad academic code ("you get what you pay for") now impressive

Alternative title: "instant gratification quantitative deconvolution fluorescence microscopy".

http://works.bepress.com/gmcnamara/55/

Please see "74"

http://works.bepress.com/gmcnamara/74/

for 32-bit images from this project (bepress file size limitation prevented me from including them in this ZIP archive).

//

Summary: Deconvolution microscopy has historically been painfully slow. The early vendors were:

- Scanalytics (Carrington and Fay), commercialized to try to sell expensive, specialized array processors made by CSPI (the CSPI box likely had less computing power than a first gen smartphone).

- Applied Precision (Sedat and Agard).

- AutoQuant (now part of Media Cybernetics.

(Note: I have not looked up when SVI, markers of Huygens, started operations - they were also a relatively early deconvolution company).

- Free (you get what you pay for:) various ImageJ plugins were both extremely painfully slow and had tedious settings (pre-2013 and currently for the plugins standard in the Fiji ImageJ distribution).

Marc Bruce & Butte's GPU deconvolution (Stanford University) was an eye opener in using a NVidia GPU card to greatly speed things up, but unfortunately had an "anti-user" interface (partly because ImageJ features bad user interfaces, partly because ImageJ is Java-based, and has bad user interface). Marc Bruce graduated (PhD), the code and he spun out to www.microvolution.com and Stanford University 'pulled the plug' on the academic version -- just as well since the academic version had some issues (this "55" ZIP file includes the old text describing some of these issues - you can call Marc Bruce to ask him for the more academic limitations). Now in mid-2015, the Microvolution software kicks butt (no, not Manish's). I am using it with an NVidia TITAN card (6 GByte RAM), purchased 12/2013. Now (6/2015) the TITAN Z card (12 Gbyte RAM) performance is 5.3 vs original TITAN is 3.5 (performance scores from NVidia). The next gen NVidia cards are likely to double the performance of "Z" in the next 12 months. Marc also implemented multi-GPU, so if your PC can run it, you could double performance (for datasets that are parallelizable enough) by running "ZZ", or triple with "ZZZ". Hopefully NVidia will come out with "O" and "W" models, so we can configure a workstation with "WOW".

//

Article comments

Data is not subject to copyright. If you are able to deconvolve (or otherwise image process and improve) the raw data, please do so, post your methods and results on the Internet and let me know where to find them and you.

Abstract

Update 20150622Mon and 20150623Tue:

This page and download

http://works.bepress.com/gmcnamara/55/

now has results using the www.Microvolution.com commercial GPU deconvolution software. Compared to the free (“you get what you pay for”) Stanford University program, Bruce and Butte 2013,

https://www.osapublishing.org/oe/fulltext.cfm?uri=oe-21-4-4766&id=249375

which was effectively limited to ~1024x1024 pixels, I now have access (and hope to go from purchase requisition to purchase order to permanent license, “soon”) to the Microvolution software that:

1. full 2048x2048 pixels (XY), so far I have used 32 plane Z-series.

2. faster than the free version on the same NVidia TITAN GPU graphics card.

3. more options ... even the most quantitative version is very fast.

4. Works with both widefield and confocal 3D Z-series, and with 2D images.

Note: How much to deconvolve is up to the user. I included both 10 and 100 iterations.

I have updated the ZIP file with:

* the original 16-plane Z-series

* 10 iterations, 16-bit … processed in 7.6 seconds (Fiji ImageJ).

* 100 iterations, 16-bit … processed in 43 seconds (Fiji ImageJ).

and left in the Stanford University 16-bit files. Bepress has a file size limit for uploads.

And kept the old Stanford University GPU deconvolutions, buth full size and cropped (TL = top left corner).

The Microvolution deconvolution settings included:

Numeric offset 200 (the lowest intensity pixel in the stack was 211).

Regularization = on.

Vectorial PSF = on.

XY pixel size 158 nm

Z step size 300 nm (Leica DMI6000).

600 nm emission wavelength (SOLA LED lamp, Leica TX2 filter cube).

NA 1.4 objective lens

RI 1.515 oil

RI 1.5 mounting medium.

http://works.bepress.com/gmcnamara/55