This program is using a series of labelling and morphological operations to extract the Lung volume intensity image from chest CT scans. It was tested with the data from the LIDC-IDRI (Lung Image Database Consortium) project and depends on ITK/cmake.
After the initial step of extracting the intensities of the lungs and airways the algorithm attempts to separate the two lungs and the airways.
This tool depends on itk and cmake. As a simple way to document (and build) the project file a Dockerfile is provided. It should be sufficient to build the docker container once
docker build -t lungsegmentation -f Dockerfile .
and to run it using
> docker run --rm -it lungsegmentation
Usage : ./LungSegmentation
System tags:
[ -v ] or [ -h ]
= List options in short format
[ -V ] or [ -H ]
= List options in long format
[ -vxml ] or [ -hxml ] or [ -exportXML ]
= List options in xml format for BatchMake
[ --xml ]
= List options in xml format for Slicer
[ -vgad ] or [ -hgad ] or [ -exportGAD ]
= List options in Grid Application Description format
[ -version ]
= return the version number
[ -date ]
= return the cvs checkout date
Command tags:
[ -n [ seriesname ] ]
= Select series by series name (if more than one series is present).
[ -b < labelfieldfilename > ]
= Save the label field as a nifty file in the current directory
[ -u < niftyfilename > ]
= Save the corrected dataset as a nifty image to the current directory
[ -V ]
= Print more verbose output
Command fields:
< indir >
= Directory with input DICOM image series.
< outdir >
= Directory for output DICOM image series.
In order to provide the data for processing the docker call should also include a '-v' to mount your data directory inside the docker container. Here an example call that assumes you have a 'data/folder_with_dicom' directory in your current directory. The following call will save the output in the same folder.
docker run --rm -it -v `pwd`/data:/data lungsegmentation /data/folder_with_dicom /data/folder_with_dicom_segmentedAdjust the CMakeLists.txt file with the path for your itk version.
cmake -DCMAKE_BUILD_TYPE=Debug .
makeThe program FakeLungVolumes generates artificial vessel volumes. As those are generated by a noise smoothing process the data is artificial and only resembles vessels. Generation of such volumes is fast and can be tailored to resemble tissues of a given scale from micro-vascular trees to images resembling larger pulmonary vessel imaged with contrast CT.
The algorithm calculates the discrete intersection of two iso-surfaces from band-pass filtered white noise volumes. Based on the amount of band-pass filtering and the threshold for the iso-surface intersection detection blood-vessel like pattern appear.
The output volume data type is limited to unsigned short to resemble medical data usually obtained at 12bit resolution. Hounsfield units are in the range of -1024...2000 so they fit well into this output range.
To generate a 64 by 64 by 64 volume:
./FakeLungVolumes /tmp/output.nii To generate a higher resolution volume
./FakeLungVolumes -k 7 -t 0.0001 -f 1 -r 128x128x128 /tmp/output.nrrdThe generated output is artificially restricted to 12bit simulating common detector resolution. The intensity range is 0 to 4096. This does not correspond to HU but may be sufficient to generate test data for machine learning algorithms.
To visualize the vessel generation process additionally to the vessels a four-class void space segmentation can be enabled. These void spaces are defined by the sign of the two band-pass filtered white noise volumes. Here an example:
./FakeLungVolumes -k 7 -t 0.0001 -w 0.0001 -f 1 -r 128x128x128 /tmp/output.niiHere are all the options:
Option outfile is required but not defined
Command tags:
[ -r [ resolution ] ]
= Specify the resolution of the volume to be generated (in pixel as in 64x64x64).
[ -k [ kernelSize ] ]
= Specify the kernel size for the Gaussian in pixel (7).
[ -i [ iterations ] ]
= Specify the number of times the Gaussian kernels are applied (2).
[ -t [ threshold ] ]
= Specify the threshold for zero-crossing (0.0001).
[ -z [ zero ] ]
= Specify at what value the intersection should be calculated (0).
[ -f [ finalsmooth ] ]
= Specify the kernel size of a smoothing with a Gaussian at the end of the process (0).
[ -n [ additivewhitenoise ] ]
= Add some noise with "mean variance" (0, 2). Additive white noise is appropriate for simulated CT images.
[ -w [ voidspaces ] ]
= Create void spaces with a given distance away from the lines. Default is that this option is not used.
In the resulting volume 0 will be the gap space right next to each vessel (label 4095) with 1, 2, 3, 4
the values of voxel that are in void space.
[ -l [ addlesion ] ]
= Specify a lesion of a specific size (5). Requires the option VoidSpaces.
[ -d [ outputdensities ] ]
= Specify the output density values used for each segmentation ("0 1 2 3 4 2048 4096"). Requires the option VoidSpaces.
[ -m [ mask ] ]
= Specify a mask file (assumption is that the mask fits in resolution with the volume created).
[ -s [ randomseed ] ]
= Specify the value used for initialization of the random numbers (time based). The same value should produce the same fields.
[ -f ]
= Ignore existing files and force overwrite.
[ -V ]
= Print more verbose output
Command fields:
< outfile >
= Exported file name.In the above options the size of the vessel structures is defined by the size of the filter kernel (-k 7) and the number of iterations that the filter is applied to the white noise input (-i 2).
Computation time for the higher resolution volume is about 0.5seconds.
In order to explain the process geometrically the 'explain' sub-directory contains a website that performs this computation in JavaScript. By changing the number of iterations of the smoothing in 1-D, 2-D, and 3-D the different features can be visualized.
There are a number of extensions to this framework. One is that multiple vessel like structures can be generated with guaranteed properties. They will never intersect and keep a given distance from each other. Those hugging lines can be generated using the -z option to specify the crossing in conjunction with the seed option -s to make sure multiple runs of the program generate the same noise pattern. Given the same seed but different crossing levels vessel structures emerge. Here an example of the resulting 3 vessel like structures using Iso-Surface displays. This data was generated using three calls to the program:
./FakeLungVolumes -s 42 -z 0 -k 7 -t 0.0001 -f 0.4 -r 192x192x192 /tmp/output.nii
./FakeLungVolumes -s 42 -z 0.002 -k 7 -t 0.0001 -f 0.4 -r 192x192x192 /tmp/output2.nii
./FakeLungVolumes -s 42 -z -0.002 -k 7 -t 0.0001 -f 0.4 -r 192x192x192 /tmp/output3.nii
At lower resolution and in 2-D cross-section:
As a final example here a closeup of a de-novo in-silico complex tissue.
In order to generate training data several options have been added. We can create a label volume and a volume with more realistic densities - and noise (mean 0 and standard deviation 50).
./FakeLungVolumes -t 0.0001 -w 0.0001 -r 64x64x64 -l 9 -n "0 50" -d "-900 -900 -900 -900 -900 50 50" /output/output.niiIn the above example we have an ellipsoid lesion of diameter (l=) 9 (pixel) with a random aspect ratio of (0.4..1) where the second and third axis are equal. The lesion ellipsoid is by default rotated around a random axis. We are adding gaussian distributed noise as a last step with a standard deviation of 50 (and a mean of 0). And we code the different regions using density values such that (-d):
- Background intensity is set to -900
- The 4 void spaces are also set to a density of -900
- The lesion has a density of 50
- The vessels have a density of 50
Given a specific seed for the random key (-s) we generate an approximate density with some noise and a label volume with:
./FakeLungVolumes -i 2 -k 5 -s 1 -t 0.0002 -w 0.0001 -r 64x64x64 -l 5 -d "-900 -900 -900 -900 -900 50 50" -n "0 30" -f 0.5 /output/output.nii
./FakeLungVolumes -i 2 -k 5 -s 1 -t 0.0002 -w 0.0001 -r 64x64x64 -l 5 -d "0 0 0 0 0 1 0" /output/label.niiIn the above label volume only the lesion (size 5) will have a label value of 1. Generating larger amounts of test data for deep learning:
mkdir /output
mkdir /label
for i in `seq 0 100000`; do
# pick a random size for the lesion aspect ratio and rotation is always random
sizes[0]=3; sizes[1]=5; sizes[2]=7; sizes[3]=9; idx=$(($RANDOM % 4)); rz=${sizes[$idx]};
# generate filenames with leading zeros
case=`printf '%06d' $i`
# write a fake-CT with a random seed of i
./FakeLungVolumes -i 2 -k 5 -s $i -t 0.0002 -w 0.0001 -r 64x64x64 -l $rz -d "-900 -900 -900 -900 -900 50 50" -n "0 30" -f 0.5 /output/${case}.nii.gz
# write the corresponding label file (0 - background, 1 - lesion)
./FakeLungVolumes -i 2 -k 5 -s $i -t 0.0002 -w 0.0001 -r 64x64x64 -l $rz -d "0 0 0 0 0 1 0" /label/${case}.nii.gz
done