White Matter, Ventricle and Subarachnoid Space Segmentation of CT scans, with Normal Pressure Hydrocephalus Predicton

Code implementing the algorithm described in the paper

Zhang et. al., Fully Automated Volumetric Classification in CT Scans for Diagnosis and Analysis of Normal Pressure Hydrocephalus. https://arxiv.org/abs/1901.09088

Requirements

To download all prerequisites, in the terminal type

pip install -r requirements.txt

In order to use the morphological chan-vese model, FSL is needed to run this code. Please go to the FSL website at https://fsl.fmrib.ox.ac.uk/fsl/fslwiki to download their software.

Note that there is a current known bug to using FSL which requires the user to manually install libopenblas.

The unet model will not require FSL. However, it requires a large file, unet_ce_hard_per_im_s8841_all/model_last.tar, which can be downloaded manually on the github website or using git-lfs.

The code has been tested only on python version 3.6.

Usage

All of the CT scans must be in compressed nifti (.nii.gz) format, in a folder named 'Scans'. Please note the directory containing the 'Scans' folder and use that as an argument --directory=</path/to/directory> when running the code. The main function nph_prediction.py takes the following arguments from the command line:

directory --directory=</path/to/directory>,

seg_model --seg_model=<model>, where <model> can be unet or mcv,

--parallel, which is True if typed and False otherwise, and

--gpu, which is True if typed and False otherwise.

To run with default settings (recommended), type

python3 nph_prediction.py

For help on the settings, type

python3 nph_prediction.py -h

The runtime is approximately 10 minutes per image.

The output will be a .csv file to <folder> with the name of the nifti image file, the white matter, ventricle, and subarachnoid space volumes, and another .csv file with the recommendation of 'normal' or 'possible NPH'.

An example CT scan has been included in the Scans directory.

This is an open source image from http://headctstudy.qure.ai/dataset, accompanying paper at:

Sasank Chilamkurthy et al. Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study. DOI:https://doi.org/10.1016/S0140-6736(18)31645-3.

Authors

Angela Zhang

Contributors

Po-yu Kao, CT registration, UNet pretraining

Austin Mcever, testing

Acknowledgements

Thanks to the Vision Research Laboratory at the University of California, Santa Barbara; Dr. Ashu Shelat at the Cottage Hospital of Santa Barbara; and Dr. Jeff Chen at the University of California, Irvine for their help and support.