Lung segmentation and statistical analisys
The main purpose of the code in this repo is to calculate the percentage of lung volume where the perfusion is lower than a user-defined threshold. In order to reach this goal a number of scripts have been implemented. They permit to:
- divide DICOM study folders into DICOM series folders
- perform lungs segmentation (thanks to lungmask)
- isolate the lungs pixels values and report them into a .csv file, divided in left and right lung
- generate histograms from .csv file, computing the ratio between pixels lower / higher than the threshold
Moreover, a set of side scripts have been implemented to further analysis and/or speed up the workflow:
- images registration
- folders unzip
In particular, at the moment the best results are obtained using images with contrast medium. The future aim is to find a way to obtain the same results without the use of contrast medium.
- Python3
- Pipenv
- Git
pipenv install
That's all!
- Install conda [don't replace default python, don't add anything to the environment variables]
- Install pytorch with
conda install pytorch torchvision cpuonly -c pytorch - Install lungmask (see its readme)
- Install needed packages (see Pipfile) with
conda installorpip install
eg:
conda install -c conda-forge pydicom
conda install -c simpleitk simpleitk
conda install -c anaconda scikit-image
If you get this error: numpy.ndarray size changed, may indicate binary incompatibility. Expected 88 from C header, got 80 from PyObject, just try to uninstall and reinstall numpy.
Run this to subdivide series in the main folder:
python utils.py --organize path/to/dicom/study/folder
To run the complete workflow (lungs extraction & stats analysis) on a folder that contains a single dicom serie, run the following command:
python complete_workflow.py Options:
--dicomdir : followed by the path to the input folder
--outdir : followed by the path to the output folder
[optional] --thresholds : followed by the three thresholds divided by spaces (lower_bound, threshold, upper_bound), eg -1000 -920 -770 [optional] --ignore_high_threshold : if the label 3 (over high threshold) volume should be included in lungs total volume [default false]
Run this to segment lungs and extract values to .csv file:
python run.py --dicomdir path/to/dicom/series/folder
Options:
--thresholds [lower_bound, threshold, upper_bound] : use this to set custom thresholds (must be same as next step)
--nrrddir path/to/nrrd/file : use instead of --dicomdir to read a .nrrd file as input
--outfolder path/to/output/folder : specify output folder (must exist), default ./output/
--force_cpu : set this to force prediction with cpu (slower), not needed if no gpu is available
--use_mask path/to/lung/mask : use this to pass a previously generated mask and skip segmentation
The output consist in a nrrd file (lung_mask_palette.nrrd) and output.csv file. Note: the input dicom serie is converted in .nrrd file in the ./temp/ directory, to be used as input to segmentation neural net.
Run this to generate histograms and compute low perfusion ratio:
python utils.py --examine path/to/csv
Options:
--thresholds [lower_bound, threshold, upper_bound] : use this to set custom thresholds (must be same as previous step)
Histogram is shown and saved in the same directory of the input file, as histogram.png and histogram.csv.
At the end of run.py script execution, the following files are produced:
output/lungs.nrrd: contains lungs extraction (everything else is set to a fixed value of -1000)output/lung_mask_palette.nrrd: contains lungs mask that shows low perfusion voxels (set to a fixed value of 10) and high perfusion voxels (set to a fixed value of 20)output/output.csv: the same as lung_mask_palette but in a tabular form
At the end of utils.py script execution, the following files are produced:
output/histogram.png: the resulting histograms & ratio low/high pefusionoutput/histogram.csv: the same data but in a tabular form
use registration.py to register images. It accepth both dicom dirs or nrrd files. Uses elastix via external script runRegistration.sh.
You can contact us at: