Teem: nrrd: Visible Female Color Data Processing: Image Differences


teem	/	nrrd	/	Visible Female	Image Differences

As yet another way of looking at the brightness variation problem, per-pixel image differences are also interesting. They have less ability to characterize the brightness differences than do the joint histograms, but they also provide some contextual clues about the nature of the brightness variations. In the previous pages on the gray scale strips and joint histograms, four locations in the head were chosen for their evidence (and non-evidence) of the brightness problem. This page looks at successive pairs of images from those four locations, in conjuction with closer inspection of the axis histograms.

cat > cmap2.txt
0 0 0
0 0 1
1 0 0
1 1 0
1 1 1
(Ctrl-D)
alias sliceprep "unu make -i - -t uchar -s 2048 1216 3 -e raw | unu permute -p 2 0 1 | unu crop -min 0 260 30 -max M 1820 M"
alias shrink "unu resample -s = x0.2 x0.2 -k cubic:0,0.5"
alias cquant "unu quantize -b 8 -min -30 -max 30"
alias nquant "unu project -a 0 -m L2 | unu rmap -m cmap2.txt -min 0 -max 60 | unu quantize -b 8"
alias histaxzoom "unu pad -min 0 -2 -2 0 -max M M+2 M+2 M -b pad -v 0 | unu permute -p 0 1 3 2 | unu reshape -s 3 42 260 | unu resample -s = x10 = -k box"

sliceprep: This crops a full-resolution slice image down to contain just the blue block and the region below it, including the Kodak card.
shrink: This downsamples from full-resolution to the image resolution used here. Note that I'm doing the differencing first, then resampling, not the other way around.
cquant: After doing image differencing, I want to re-quantize to 8-bits, but want to set the min and max explicitly to be centered around zero, so that the difference of identical pixels will be mapped to (128,128,128).
nquant: After doing image differencing, this takes the L2 norm of the component-wise difference, and then colormaps this through cmap2.txt
histaxzoom: Its possible to do image manipulation with unu, but its not pretty or intuitive, and this is a good example of that. In conjuction with the unu join and unu crop on the per-component axis histograms, this adds a black border around the selected region, re-arranges the images side-by-side, then upsamples in X, so as to "zoom" in on axis histograms of 5 slices on either side of the transition of interest.

Since many images are involved, the sets of difference images are organized into sub-pages.

First step discontinuity: This is the discontinuity that goes through the corpus callosum, the streak most visible in the XZ and YZ projections of the cube of brain.
Second step discontinuity: This is the discontinuity that goes through the tops of the eyes.
Isolated dark slice: This is the single dark slice that goes through the bridge of the nose.
The control: This region was deemed to have no problematic brightness variations.

Although it is a stretch to make strong conclusions based on a small number of slices, these difference images imply that there are at least two very different causes of brightness variations. In the first two cases (the step discontuities) the change was exactly coincident with the Kodak card being moved. Perhaps some aspect of the machinery was reset at these slices, including a change in the lighting. It is in these cases where the Kodak card did successfully capture the brightness variation that occurs inside the head.

In the third case, the Kodak card has not moved, and the pattern on the difference images suggest that something changed on the surface itself, rather than the lighting, which would explain why the Kodak card completely failed to capture the brightness variation visible inside the head. Based on this, I did difference images for another set of images around an isolated dark slice, the one that goes through the nose, but closer to the tip than the previous dark slice. Slice indices and names are:

459: avf1154a.raw.Z
460: avf1154b.raw.Z
461: avf1154c.raw.Z
462: avf1155a.raw.Z
463: avf1155b.raw.Z

unu join -i histax{R,G,B}.ppm -a 3 | unu crop -min 0 456 0 0 -max M m+9 M M \
  | histaxzoom | topng doc/hist5.png
zcat data/avf1154a.raw.Z | sliceprep -o 0.ppm
zcat data/avf1154b.raw.Z | sliceprep -o 1.ppm
zcat data/avf1154c.raw.Z | sliceprep -o 2.ppm
zcat data/avf1155a.raw.Z | sliceprep -o 3.ppm
zcat data/avf1155b.raw.Z | sliceprep -o 4.ppm
unu 2op - {0,1}.ppm -t int | cquant | shrink | topng doc/diff5CA.png
unu 2op - {1,2}.ppm -t int | cquant | shrink | topng doc/diff5CB.png
unu 2op - {2,3}.ppm -t int | cquant | shrink | topng doc/diff5CC.png
unu 2op - {3,4}.ppm -t int | cquant | shrink | topng doc/diff5CD.png
unu 2op - {0,1}.ppm -t int | nquant | shrink | topng doc/diff5NA.png
unu 2op - {1,2}.ppm -t int | nquant | shrink | topng doc/diff5NB.png
unu 2op - {2,3}.ppm -t int | nquant | shrink | topng doc/diff5NC.png
unu 2op - {3,4}.ppm -t int | nquant | shrink | topng doc/diff5ND.png


`hist5.png`: Segments of the red, green, and blue axis histograms, for Z slices 456 to 465.

`diff5CA.png`	`diff5CB.png`	`diff5CC.png`	`diff5CD.png`

`diff5NA.png`	`diff5NB.png`	`diff5NC.png`	`diff5ND.png`

As with the previous isolated dark slice, the Kodak card indicates no change, and, greater difference magnitudes on the slice were visible to the sides of the head. The colormapping of the L2 difference norm shows that the interior of the head also has higher difference values; there are more yellow pixels in the back of the head in diff4NB.png and diff4NC.png.