[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [ccp4bb]: Improving data processing with Gaussian blur

Title: Message
It is also known that Photoshop is the most
effective density modification program available....
It is believed that together with combinatorial
data generation and phases from predicion models it
will replace CCP4 with the next few years.
Adobe Xtallator...
PS: dibbs on that patent.
-----Original Message-----
From: owner-ccp4bb@dl.ac.uk [mailto:owner-ccp4bb@dl.ac.uk] On Behalf Of Mischa Machius
Sent: Tuesday, April 01, 2003 7:43 AM
To: CCP4 Bulletin Board
Cc: Mischa Machius
Subject: [ccp4bb]: Improving data processing with Gaussian blur

Hi y'all,

We have been looking for a program to print diffraction images and found
Photoshop to be very useful for a variety of CCD formats. When playing
around with some options, we stumbled upon an interesting effect with
images from the SBC2 CCD detecor (APS beamline ID19): When using a
Gaussian blur with a radius of 2.3, data processing and structure
refinement improved significantly. R syms for typical datasets get better
by about 2-3% overall, but the I/sigma(I) decreases somewhat. Refinement
with CNS yielded R factors between 1-4% better (0.5-2% in Rfree).
The filter has to applied in CMYK space, gray scale and RGB don't
work. Heated discussions are going on here why this might be.

The effect is similar to binning followed by lossy compression and
uncompression. The procedure seems to even out spot background.
An alternative explanation is that the Gaussian blur acts like data
sharpening by applying a B factor. Interestingly, the radius for the
Gaussian blur needs to be between 2.1 and 2.5, otherwise data processing
gets really bad.

A similar effect can be achieved by the "Pointillize" command in the
"Pixelate" submenu. Here, the cell size should be between 7 and 13% for
the procedure to work, but the chi^2 values in Denzo will be artificially
lowered while the B factors of refined models in CNS will be too high by a
constant factor (compared to non-treated data). However, the electron
density looks qualitatively better than the one obtained with the Gaussian
blur and significantly better than without blurring the images.

>From now on, we will routinely do a pre-processing of our images with
Photoshop. However, it doesn't work in all cases: image plate data did not
improve, whereas CCD data from beamlines with beams that are not entirelyhomogeneous did benefit a lot.

Our results strongly suggest to include Gaussian blur and pixelating
protocols in current data processing software, although going through
Photoshop is not such a big deal, because it can be automated.

Any thoughts?


Mischa Machius, PhD
Assistant Professor
UT Southwestern Medical Center at Dallas
5323 Harry Hines Blvd.; Y4.330A
Dallas, TX 75390-9038; U.S.A.
Tel: +1 214 648 9760
Fax: +1 214 648 8954