*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy*****
GPU-based 3D Deconvolution update: TITAN card enables FLASH4.0 full
sensor Z-series
Bruce, Butte 2013, Real-time GPU-based 3D Deconvolution, Optics Express:
I've had their ImageJ plugin since publication (free to academic users).
With my old GPU card (Quadro 2000) I could process a 512x512x32 plane
stack with 100 iterations (far nicer than typical default that ends
around 30) in ~7 seconds (faster than it takes to type in all the
parameters!) but was unable to process Hamamatsu FLASH4.0 sCMOS
2,048x2048 pixel images of any number of planes. We just installed a
NVidia (EVGA) GeForce GTX TITAN card (6 gigabytes ram, 2880 processors,
faster clock, faster data transfer, list price around $1050, UT MDACC
might have paid a bit less than internet list price), now takes 1.6
seconds to do 100 iterations of 512x512x32 plane, and 17 second for
2048x2048x32 planes (single channel ... can also operate with "RGB"
16-bit images). After exchanging emails with Manish Butte I now know to
use power of 2 number of Z-planes (and knew to use power of 2 XY pixels).
http://www.opticsinfobase.org/oe/fulltext.cfm?uri=oe-21-4-4766http://tcell.stanford.edu/software.htmlI am not claiming that B&B GPU deconvolution produces the best possible
images - in fact, SVI has sent me very nice results from the "31"
dataset (url below) and I have their new Huygens 4.5 on trial this week
(thanks Remko and SVI!). I do like the B&B GPU speed.
I often acquire with 500 millisecond exposure times (SOLA light engine,
Leica 63x/1.4NA objective lens, Leica filter cubes, FLASH4.0,
MetaMorph), so 32 planes is 16 seconds, so GPU time is about the same
(ignoring typing in parameters). Typical specimens recently have been
similar to those posted at
http://stellarisfish.smugmug.com/I have some raw FISH data posted at
http://works.bepress.com/gmcnamara/31/http://works.bepress.com/gmcnamara/32/ (part 1 of 2)
http://works.bepress.com/gmcnamara/33/ (part 2 of 2)
enjoy,
George
p.s. Acquisition time I used depends on experiment ... With the same
microscope I have acquired mitochondria targeted YFP in live human
T-cells with 1 millisecond exposure time per plane (1/10th chip chip ROI
or smaller ... by the way, for non-square ROIs on the FLASH, orientation
affects speed, fortunately 512x512 pixels is 1/16th area and square),
using nearly the entire dynamic range of the FLASH4.0 (16-bit output).
Imaging conditions are not perfect because the cells were in aqueous
solution, not glued down. Very little photobleaching because we had
Marker Gene Tech's Opti-Klear in the media (I haven't tested Evrogen's
DMEMgfp and don't know if they work the same way). I don't expect
perfection in any deconvolution with spherical cells in aqueous buffer
because the cells themselves act as crystal balls, not infinitely thin
layer right at the coverglass.
--
George McNamara, Ph.D.
Single Cells Analyst
L.J.N. Cooper Lab
University of Texas M.D. Anderson Cancer Center
Houston, TX 77054
Tattletales
http://works.bepress.com/gmcnamara/26/
Locked
