Examples
CT Pre-processing
Flat field correction
To remove fixed pattern noise that stems from the optical system caused by imperfections in the scintillator screen or an inhomogeneous beam and thermal noise from the detector sensor, you can use flat field correction. This assumes that you have a set of dark fields acquired with the shutter closed, a set of flat fields acquired without the sample in the beam and the projections with samples. If the beam intensity shifts over time it can be beneficial to acquire flat fields before and after the projections and interpolate between them.
In the simplest case you connect the projection stream to input 0, the dark
field to input 1 and the flat field to input 2 of
flat-field-correct:
ufo-launch \
[ \
read path=projections*.tif, \
read path=dark.tif, \
read path=flat.tif \
] ! \
flat-field-correct !
write filename=corrected-%05i.tif
If you have a stream of flats and darks you have to reduce them either by
connection them to average or stack →
flatten with the mode set to median. Suppose, we want to
average the darks and remove extreme outliers from the flats, we would call
ufo-launch \
[ \
read path=projections*.tif, \
read path=darks/ ! average, \
read path=flats/ ! stack number=11 ! flatten mode=median \
] ! \
flat-field-correct !
write filename=corrected-%05i.tif
If you have to interpolate between the flats taken before and after the sample
scan, you would connect the first flat to input 0 and the second to input 1 of
interpolate and set the number property to the number of
expected projections:
ufo-launch \
[ \
read path=projections*.tif, \
read path=darks/ ! average, \
[ \
read path=flat-before.tif, \
read path=flat-after.tif \
] ! interpolate number=2000
] ! \
flat-field-correct !
write filename=corrected-%05i.tif
If you want to avoid the automatic absorption correction you have to set
absorption-correct to FALSE and if you want to ignore NaN and Inf values in
the data, set fix-nan-and-inf to FALSE.
Sinograms
The reconstruction pipelines presented in the following section assume sinograms
as input in order to parallelize along slices. To transpose a stream of
(corrected) projections connect it to transpose-projections and
set number to the number of expected projections. Note, that the
transposition happens in main memory and thus may exhaust your system resources
for a larger number of big projections. For example, to transpose 2048
projections, each at a size of 2048 by 2048 pixels requires 32 GB of RAM.
CT Reconstruction
Filtered backprojection
To reconstruct from sinograms using the analytical filtered backproject method
[KaSl01], you have to feed the sinograms into fft →
filter → ifft → backproject to obtain
slices one by one:
ufo-launch \
dummy-data width=$DETECTOR_WIDTH height=$N_PROJECTIONS number=$N_SLICES ! \
fft dimensions=1 ! \
filter ! \
ifft dimensions=! ! \
backproject axis-pos=$AXIS ! \
null
Direct Fourier inversion
In this example we use the Fourier slice theorem to obtain slices directly from
projection data [KaSl01] and use a sinc kernel to interpolate in the Fourier
space. To reconstruct, you have to feed the sinograms into zeropad
→ fft → dfi-sinc → swap-quadrants →
ifft → swap-quadrants
ufo-launch \
dummy-data width=$DETECTOR_WIDTH height=$N_PROJECTIONS number=$N_SLICES ! \
zeropad center-of-rotation=$AXIS ! \
fft dimensions=1 auto-zeropadding=0 ! \
dfi-sinc ! \
swap-quadrants ! \
ifft dimensions=2 ! \
swap-quadrants ! \
null
Data distribution
To distribute data in a compute network you can use the zmq-pub
sink and zmq-sub generator. For example, to read data on machine A
and store it on machine B, you would run
ufo-launch read path=/data ! zmq-pub
on machine A and
ufo-launch zmq-sub address=tcp://hostname-of-machine-a ! write
on machine B. Note that by default zmq-pub publishes data as soon
as it receives it, thus some of the data will get lost if the
zmq-sub is run after zmq-pub. You can prevent this
by telling the zmq-pub task to wait for a certain number of
subscribers to subscribe:
ufo-launch read path=/data ! zmq-pub expected-subscribers=1
References