GPU deconvolution on TITAN card follow-up

[George McNamara]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
*****

GPU deconvolution on TITAN card follow-up ...

NVidia TITAN card needs to be in PCI Express (version 2) x16 slot and
the Hamamatsu FLASH4.0's CameraLink card (actve silica) needs to be in
x8 or x16 ... the FLASH4.0 was originally installed in an x4 slot, which
usually worked with Auadro 2000 cards but not with the new card.
My thanks to Hamamatsu tech support for identifying the CameraLink card
was installed in a sub-optimal slot and our local I.T. support (Thai)
for re-locating the cards in the T7500. We acquired and deconvolved
2048x2048, 32 plane (100x100 nm pixels, 200 nm Z steps) of single
molecules Stellaris FISH probes (see http://stellarisfish.smugmug.com 
for Biosearch's image gallery) in 16 seconds (500 ms exposure times) and
deconvolved in 18 seconds (Quasar 670 dye, Cy5 filter cube, SOLA light
engine, not counting opening Fiji ImageJ, the file, the B&B
deconvolution dialog typing in parameters and clicking run).
Manish Butte's web site and article
http://tcell.stanford.edu/software.html
http://www.opticsinfobase.org/oe/abstract.cfm?uri=oe-21-4-4766


    Real-time GPU-based 3D Deconvolution - Optics Express (open access)


With respect to Optics Infobase, Gary Brooker recently published new
hardware for FINCH - looking pretty cool
http://www.opticsinfobase.org/ol/abstract.cfm?URI=ol-38-24-5264
We report a new optical arrangement that creates high-efficiency,
high-quality Fresnel incoherent correlation holography (FINCH) holograms
using polarization sensitive transmission liquid crystal gradient index
(TLCGRIN) diffractive lenses. In contrast, current universal practice in
the field employs a reflective spatial light modulator (SLM) to separate
sample and reference beams. Polarization sensitive TLCGRIN lenses enable
a straight optical path, have >90% transmission efficiency, are not
pixilated, and are free of many limitations of reflective SLM devices.
For each sample point, two spherical beams created by a glass lens in
combination with a polarization sensitive TLCGRIN lens interfere and
create a hologram and resultant super resolution image.
(unfortunately Optics Letters is not open access).


Why I am interested in GPU (or Intel Phi coprocessor or Cloud)
deconvolution, FINCH, etc? - In 1995, Niswender et al
(http://www.ncbi.nlm.nih.gov/pubmed/?term=8537958) published (pre-EGFP)
that,
"In HeLa cells, the cytoplasmic GFP concentration must be greater than
approximately 1 microM to allow quantifiable discrimination over
autofluorescence. However, lower expression levels may be detectable if
GFP is targeted to discrete subcellular compartments, such as the plasma
membrane, organelles or nucleus."
Most researchers just spew FP-fusions throughout the entire cytoplasm
(and/or nucleoplasm, depending on size and/or nuclear localization
signal). localized FPs are much nicer to work with - and make
diffraction limited spots ideal for spatial deconvolution.
I am excited by our opportunities to localize fluorescent protein
biosensors (http://works.bepress.com/gmcnamara/26/). Latest published FP
biosensor reports on central metabolism (my UIUC grad school biochem
prof, Stephen Sligar, should be amazed that I have any interest in
biochemistry):
http://www.ncbi.nlm.nih.gov/pubmed/?term=pkm2+biosensors
and Gary Yellen recently published a High dynamic Range version of their
ATP:ADP ratio biosensor, http://www.ncbi.nlm.nih.gov/pubmed/?term=24096541
and Jin Zhang et al edited a new MMB on FP biosensors (6 Mb pdf download
for places with online access)
http://link.springer.com/book/10.1007/978-1-62703-622-1
Of course I am also excited by our being able to image and count single
molecules RNA by FISH -- and maybe single molecules in live cells with
molecular beacons. I have also encouraged the EMD Millipore SmartFlare
R&D/sales that they should come up with a way(s) to localize the
fluorescent "leaving group" in that product line.

Enjoy,

George

--



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/

[Brian Northan]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
*****

George

The reported times are really good.  I looked at the images and noticed
that (as expected from experiment description) there was some spherical
aberration.  Are there any settings in the GPU algorithm PSF generator to
deal with that??

I took a quick shot at PSF generation and Deconvolution using other ImageJ
plugins (PSF Generator and Deconvolution Lab from EPFL (
http://bigwww.epfl.ch/algorithms/deconvolutionlab/)).  The PSF is still an
approximation, but considers SA using these guidelines (
http://www.ncbi.nlm.nih.gov/pubmed/21118213).

Results and description can be found here...

http://truenorth-ia.com/blog/


On Tue, Dec 17, 2013 at 9:31 PM, George McNamara
<[hidden email]>wrote:

> *****
> To join, leave or search the confocal microscopy listserv, go to:
> http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
> *****
>
> GPU deconvolution on TITAN card follow-up ...
>
> NVidia TITAN card needs to be in PCI Express (version 2) x16 slot and the
> Hamamatsu FLASH4.0's CameraLink card (actve silica) needs to be in x8 or
> x16 ... the FLASH4.0 was originally installed in an x4 slot, which usually
> worked with Auadro 2000 cards but not with the new card.
> My thanks to Hamamatsu tech support for identifying the CameraLink card
> was installed in a sub-optimal slot and our local I.T. support (Thai) for
> re-locating the cards in the T7500. We acquired and deconvolved 2048x2048,
> 32 plane (100x100 nm pixels, 200 nm Z steps) of single molecules Stellaris
> FISH probes (see http://stellarisfish.smugmug.com for Biosearch's image
> gallery) in 16 seconds (500 ms exposure times) and deconvolved in 18
> seconds (Quasar 670 dye, Cy5 filter cube, SOLA light engine, not counting
> opening Fiji ImageJ, the file, the B&B deconvolution dialog typing in
> parameters and clicking run).
> Manish Butte's web site and article
> http://tcell.stanford.edu/software.html
> http://www.opticsinfobase.org/oe/abstract.cfm?uri=oe-21-4-4766
>
>
>    Real-time GPU-based 3D Deconvolution - Optics Express (open access)
>
>
> With respect to Optics Infobase, Gary Brooker recently published new
> hardware for FINCH - looking pretty cool
> http://www.opticsinfobase.org/ol/abstract.cfm?URI=ol-38-24-5264
> We report a new optical arrangement that creates high-efficiency,
> high-quality Fresnel incoherent correlation holography (FINCH) holograms
> using polarization sensitive transmission liquid crystal gradient index
> (TLCGRIN) diffractive lenses. In contrast, current universal practice in
> the field employs a reflective spatial light modulator (SLM) to separate
> sample and reference beams. Polarization sensitive TLCGRIN lenses enable a
> straight optical path, have >90% transmission efficiency, are not
> pixilated, and are free of many limitations of reflective SLM devices. For
> each sample point, two spherical beams created by a glass lens in
> combination with a polarization sensitive TLCGRIN lens interfere and create
> a hologram and resultant super resolution image.
> (unfortunately Optics Letters is not open access).
>
>
> Why I am interested in GPU (or Intel Phi coprocessor or Cloud)
> deconvolution, FINCH, etc? - In 1995, Niswender et al (
> http://www.ncbi.nlm.nih.gov/pubmed/?term=8537958) published (pre-EGFP)
> that,
> "In HeLa cells, the cytoplasmic GFP concentration must be greater than
> approximately 1 microM to allow quantifiable discrimination over
> autofluorescence. However, lower expression levels may be detectable if GFP
> is targeted to discrete subcellular compartments, such as the plasma
> membrane, organelles or nucleus."
> Most researchers just spew FP-fusions throughout the entire cytoplasm
> (and/or nucleoplasm, depending on size and/or nuclear localization signal).
> localized FPs are much nicer to work with - and make diffraction limited
> spots ideal for spatial deconvolution.
> I am excited by our opportunities to localize fluorescent protein
> biosensors (http://works.bepress.com/gmcnamara/26/). Latest published FP
> biosensor reports on central metabolism (my UIUC grad school biochem prof,
> Stephen Sligar, should be amazed that I have any interest in biochemistry):
> http://www.ncbi.nlm.nih.gov/pubmed/?term=pkm2+biosensors
> and Gary Yellen recently published a High dynamic Range version of their
> ATP:ADP ratio biosensor, http://www.ncbi.nlm.nih.gov/pubmed/?term=24096541
> and Jin Zhang et al edited a new MMB on FP biosensors (6 Mb pdf download
> for places with online access) http://link.springer.com/book/
> 10.1007/978-1-62703-622-1
> Of course I am also excited by our being able to image and count single
> molecules RNA by FISH -- and maybe single molecules in live cells with
> molecular beacons. I have also encouraged the EMD Millipore SmartFlare
> R&D/sales that they should come up with a way(s) to localize the
> fluorescent "leaving group" in that product line.
>
> Enjoy,
>
> George
>
> --
>
>
>
> 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/
>

[George McNamara]

*****
To join, leave or search the confocal microscopy listserv, go to:
http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
*****

Hi Brian,

There are no GPU deconvolution software settings for spherical aberration.

This could be dealt with by matching the refractive index of the
mounting medium to that of the glass and immersion medium ... or at
least matching the two media. Ther is also the issue that mounting media
are not necessarily well miscible with aqueous media. Stan Vitha has
posted on the listserv, and published, on transitioning from aqueous to
2,2'-thiodiethanol (TDE), re:

2,2'-thiodiethanol: a new water soluble mounting medium for high
resolution optical microscopy. <http://www.ncbi.nlm.nih.gov/pubmed/17131355>

*Staudt* T, Lang MC, Medda R, Engelhardt J, *Hell* SW.

Microsc Res Tech. 2007 Jan;70(1):1-9.

PMID:
    17131355



It would be great if the other ImageJ plugins worked reliably and were
fast - the ones I've tried recently have been neither.

The GPU deconvolution has an option to use its own PSF generator or let
the user provide a file - I leave it checked on. Makes more sense to me
to optimize specimen preparation as in 2nd paragraph above.

A key to the success of using the TITAN card is that it has 6 Gb ram,
whereas the card I almost ordered has 3 Gb ram. I ordered the TITAN
after an exchange of emails with Prof. Manish Butte and thank Manish for
answering my questions. In general I encourage both "read the
manual/publication" and communicating with the right person/people about
new equipment and software.


George


On 12/18/2013 7:00 AM, Brian Northan wrote:

> George
>
> The reported times are really good.  I looked at the images and
> noticed that (as expected from experiment description) there was some
> spherical aberration.  Are there any settings in the GPU algorithm PSF
> generator to deal with that??
>
> I took a quick shot at PSF generation and Deconvolution using other
> ImageJ plugins (PSF Generator and Deconvolution Lab from EPFL
> (http://bigwww.epfl.ch/algorithms/deconvolutionlab/)
> <http://bigwww.epfl.ch/algorithms/deconvolutionlab/%29>).  The PSF is
> still an approximation, but considers SA using these guidelines
> (http://www.ncbi.nlm.nih.gov/pubmed/21118213).
>
> Results and description can be found here...
>
> http://truenorth-ia.com/blog/
>
>
> On Tue, Dec 17, 2013 at 9:31 PM, George McNamara
> <[hidden email] <mailto:[hidden email]>> wrote:
>
>     *****
>     To join, leave or search the confocal microscopy listserv, go to:
>     http://lists.umn.edu/cgi-bin/wa?A0=confocalmicroscopy
>     *****
>
>     GPU deconvolution on TITAN card follow-up ...
>
>     NVidia TITAN card needs to be in PCI Express (version 2) x16 slot
>     and the Hamamatsu FLASH4.0's CameraLink card (actve silica) needs
>     to be in x8 or x16 ... the FLASH4.0 was originally installed in an
>     x4 slot, which usually worked with Auadro 2000 cards but not with
>     the new card.
>     My thanks to Hamamatsu tech support for identifying the CameraLink
>     card was installed in a sub-optimal slot and our local I.T.
>     support (Thai) for re-locating the cards in the T7500. We acquired
>     and deconvolved 2048x2048, 32 plane (100x100 nm pixels, 200 nm Z
>     steps) of single molecules Stellaris FISH probes (see
>     http://stellarisfish.smugmug.com for Biosearch's image gallery) in
>     16 seconds (500 ms exposure times) and deconvolved in 18 seconds
>     (Quasar 670 dye, Cy5 filter cube, SOLA light engine, not counting
>     opening Fiji ImageJ, the file, the B&B deconvolution dialog typing
>     in parameters and clicking run).
>     Manish Butte's web site and article
>     http://tcell.stanford.edu/software.html
>     http://www.opticsinfobase.org/oe/abstract.cfm?uri=oe-21-4-4766
>
>
>        Real-time GPU-based 3D Deconvolution - Optics Express (open access)
>
>
>     With respect to Optics Infobase, Gary Brooker recently published
>     new hardware for FINCH - looking pretty cool
>     http://www.opticsinfobase.org/ol/abstract.cfm?URI=ol-38-24-5264
>     We report a new optical arrangement that creates high-efficiency,
>     high-quality Fresnel incoherent correlation holography (FINCH)
>     holograms using polarization sensitive transmission liquid crystal
>     gradient index (TLCGRIN) diffractive lenses. In contrast, current
>     universal practice in the field employs a reflective spatial light
>     modulator (SLM) to separate sample and reference beams.
>     Polarization sensitive TLCGRIN lenses enable a straight optical
>     path, have >90% transmission efficiency, are not pixilated, and
>     are free of many limitations of reflective SLM devices. For each
>     sample point, two spherical beams created by a glass lens in
>     combination with a polarization sensitive TLCGRIN lens interfere
>     and create a hologram and resultant super resolution image.
>     (unfortunately Optics Letters is not open access).
>
>
>     Why I am interested in GPU (or Intel Phi coprocessor or Cloud)
>     deconvolution, FINCH, etc? - In 1995, Niswender et al
>     (http://www.ncbi.nlm.nih.gov/pubmed/?term=8537958) published
>     (pre-EGFP) that,
>     "In HeLa cells, the cytoplasmic GFP concentration must be greater
>     than approximately 1 microM to allow quantifiable discrimination
>     over autofluorescence. However, lower expression levels may be
>     detectable if GFP is targeted to discrete subcellular
>     compartments, such as the plasma membrane, organelles or nucleus."
>     Most researchers just spew FP-fusions throughout the entire
>     cytoplasm (and/or nucleoplasm, depending on size and/or nuclear
>     localization signal). localized FPs are much nicer to work with -
>     and make diffraction limited spots ideal for spatial deconvolution.
>     I am excited by our opportunities to localize fluorescent protein
>     biosensors (http://works.bepress.com/gmcnamara/26/). Latest
>     published FP biosensor reports on central metabolism (my UIUC grad
>     school biochem prof, Stephen Sligar, should be amazed that I have
>     any interest in biochemistry):
>     http://www.ncbi.nlm.nih.gov/pubmed/?term=pkm2+biosensors
>     and Gary Yellen recently published a High dynamic Range version of
>     their ATP:ADP ratio biosensor,
>     http://www.ncbi.nlm.nih.gov/pubmed/?term=24096541
>     and Jin Zhang et al edited a new MMB on FP biosensors (6 Mb pdf
>     download for places with online access)
>     http://link.springer.com/book/10.1007/978-1-62703-622-1
>     Of course I am also excited by our being able to image and count
>     single molecules RNA by FISH -- and maybe single molecules in live
>     cells with molecular beacons. I have also encouraged the EMD
>     Millipore SmartFlare R&D/sales that they should come up with a
>     way(s) to localize the fluorescent "leaving group" in that product
>     line.
>
>     Enjoy,
>
>     George
>
>     --
>
>
>
>     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/
>
>

--



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/