Code used for the CETUS Challenge, MICCAI 2014.

The Random Forest code is in C++, wrapped with Cython, with the autocontext loop performed in Python.

The training and test data can be downloaded from the challenge evaluation platform after creation of an account:

https://miccai.creatis.insa-lyon.fr/miccai/

The ground truth for the test data is not provided and the segmentation results are evaluated by uploading the corresponding meshes on the online platform.

After downloading the data, the first step is to preprocess it:

• convert the images (.mhd/.raw) and the ground truth meshes (*.vtk) to the NIfTI format;
• resample to isotropic resolution (1mm by 1mm by 1mm);
• denoise: during the challenge, non-local means denoising was used, but the corresponding Matlab code is not provided here, instead a simple median filtering is used.

The conversion can be done using the scripts convert.py and polydata2nifti.py, provided with python-irtk. This scripts/ directory must be added to your $PATH (for instance in your .bashrc). The resampling can be done with the binary resample that comes with IRTK (make sure it is on your $PATH as well (which should be the case if you installed IRTK through conda). The script convert.sh can be used to perform all the steps of this preprocessing.

Create a folder Images/ then copy all the folders Patient* into it. Similarly, create a folder Truth/ and copy all the *.vtk meshes into it. Then run convert.sh and all the data will be converted in the folder nifti/, nifti_resampled/ and denoised/ .

Install the necessary dependencies:

conda install boost opencv cython joblib
conda install -c kevin-keraudren python-irtk

Run:

cd forest && make

Fiddle with the parameters in the params dictionary, then run:

python train.py

An already trained detector is provided in forest_615_{0..5} and corresponds to the model used during the challenge.

Run:

 ./predict_all_frames.sh 30 45  # give the same ID twice to run only this subject
 ./make_submission.sh 4 # this argument is the number of autocontext iterations

Note that mcubes from IRTK is used in make_submission.sh to create a mesh from a NiFTI volume using the marching cubes algorithm, so make sure it is on your $PATH as well.

This gist can be used to take screenshots of the error meshes opened with VTK.

The script show.py can be used to make visualisations of the resulting segmentation:

The code cython code of the geodesic distance distance transform has been written by Dr. Zehan Wang as part of spatial-pbs.

Keraudren, K., Oktay, O., Shi, W., Hajnal, J.V., & Rueckert, D.: Endocardial 3D Ultrasound Segmentation using Autocontext Random Forests, CETUS Challenge, MICCAI 2014.
PDF video poster slides