Teem: nrrd: Visible Female Color Data Processing: Cube of Brain


teem	/	nrrd	/	Visible Female	Cube of Brain

One should wonder at this point if these discontinuities in intensity will ever be evident in later stages of visualization or processing. The Visible Female RGB color brain data is higher resolution (0.33 mm/pixel) than than the CT (0.48 mm/pixel) or the MRI (0.86 mm/pixel), so it would be a pity if the inter-slice brightness variations caused difficulties.

To test this, we select (from the full-resolution head images) a small cubical region inside the brain, which contains some of the discontinuities. The region to process was selected based on inspecting the xsum.png and ysum.png images above to find which Z slices were contained the brain (as aided by a mapping from Z index to slice name (slicelist.txt), created by a little C program (slicelist.c)), as well as inspecting the cropped head PPM images themselves. Once the cropped volume is created, we can do summed projections along X, Y, and Z to get a sense of what's inside.

unu join -i avf10{44,45,46,47,48,49}?.raw.Z.head.ppm \
      avf10{5,6,7,8,9}??.raw.Z.head.ppm avf110??.raw.Z.head.ppm \
      avf1110a.raw.Z.head.ppm avf1110b.raw.Z.head.ppm -a 3 \
  | unu crop -min 0 157 234 0 -max M m+199 m+199 M -o brcube.nrrd
unu project -i brcube.nrrd -a 1 -m sum \
  | unu quantize -b 8 | topng doc/brcubeYZ.png
unu project -i brcube.nrrd -a 2 -m sum \
  | unu quantize -b 8 | topng doc/brcubeXZ.png
unu project -i brcube.nrrd -a 3 -m sum \
  | unu quantize -b 8 | topng doc/brcubeXY.png

unu join ... The long complicated argument is just a way of specifying to the Unix shell that we want all head images for slices starting with avf1044a.raw.Z and ending with avf1110b.raw.Z. This is 200 slices.
unu crop ... We want a 200x200 window on each slice. The lower corner of the window is at (X,Y)=(157,234), so the lower corner of the 4-dimensional cropping is "-min 0 157 234 0", because the axis ordering is (RGB,X,Y,Z). We want to include all the colors and Z coordinates (signified by "M" with the "-max" flag), but we want to go 199 samples above the minimum in X and Y (signified by "m+199").

brcubeYZ.png brcubeXZ.png brcubeXY.png

The region selected contains nearly all of the corpus callosum, as well as a significant portion of the lateral ventricles. The banding is very visible in the YZ and XZ projections, and there is clearly a sudden change (about halfway down in XZ and YZ projections) which causes the white matter to appear with two distinct brightnesses.

Another way to see the effect of the inter-slice brightness variations is to create a derivative volume from the color volume. At each location in the RGB volume we can measure a kind of first derivative which is a 3x3 matrix. It is the sum of the outer products of the color component gradient vectors. The L2 norm of this matrix is a good indicator of the amount of local change; this metric has been used by computer vision researchers doing feature detection in color images. Teem comes with a program for doing just this, called vprobe. Actually, vprobe can measure various kind of values and derivatives in both scalar and vector volumes, but we only need it for a specific derivative measure in 3-vector data.

vprobe -i brcube.nrrd -k vector -q l2mg -k00 cubic:0,0.5 -k11 cubicd:1,0 -o mgbr.nrrd
unu project -i mgbr.nrrd -a 0 -m sum \
  | unu quantize -b 8 -min 6000 -max 100000 | topng doc/mgbrYZ.png
unu project -i mgbr.nrrd -a 1 -m sum \
  | unu quantize -b 8 -min 6000 -max 100000 | topng doc/mgbrXZ.png
unu project -i mgbr.nrrd -a 2 -m sum \
  | unu quantize -b 8 -min 6000 -max 100000 | topng doc/mgbrXY.png

unu quantize ... Giving explicit min and max bounds for the quantization over-rides the default action of using the min and max floating point value found in the input. Using the same quantization ranges for all three images makes their visual comparison more meaningful.

mgbrYZ.png mgbrXZ.png mgbrXY.png

The streaks caused by the brightness variations are clear here too, especially the one about halfway down the YZ and XZ images. The problem is that these streaks are just as bright or brighter than the gradients caused by real anatomical features, such as the boundary between white and gray matter. One could argue, though, that these axis aligned summation projections artificially accentuate the streaks caused by inter-slice brightness variations, because the streaks themselves are axis aligned. But they are clear on a single slice as well (shown alongside the projection of the same axis):

unu slice -i mgbr.nrrd -a 0 -p 43 \
  | unu quantize -b 8 -min 0 -max 800 | topng doc/mgbrX043.png

mgbrYZ.png mgbrX043.png