Chest CT registration

This repository contains the code for registration of Chest CT done with inspiratory and expiratory breath-hold CT image pairs. The dataset used is COPDGene dataset. The dataset has landmarks for all the inhale-exhale image pairs which are used to calculate the registration error.

DATA STRUCTURE

Structure of the data in our project folder

The data in the cwd() of the project must be in the following Path: cwd()/data/YOUR_DATASET.

Inside YOUR_DATASET folder one folder per patient named: copd*X*. Inside that folders all the data per patient is located with the following names changing X to a different integer per patient:

• copdX_300_iBH_xyz_r1.txt
• copdX_eBHCT.img
• copdX_iBHCT.img
• optional: copdX_300_eBH_xyz_r1.txt (groundtruth)

For this project copd1, copd2, copd3, copd4 from COPDGene were used as 'train' dataset (YOUR_DATASET).

Add at the top of the inhale landmarks .txt files the following rows:

index
300

• This is a necessary step for transformix to be able to read the inhale landmarks .txt files correctly.

Setting up the environment

• Create a conda environment

conda create -n ctreg python==3.9.13 anaconda -y && conda activate ctreg

• Install the requirements

pip install -r requirements.txt

Data preprocessing

1. Transformation of the dataset images from raw format to NIFTI format.

The initial data is in binary/raw format. The read_raw.py script assumes that the train data is in directory data/train with the patient wise .img inhale and exhale files structured as copd1,copd2,...etc. Make sure that the metadata information such as image dimensions and voxel spacings are present in copd_metadata.csv.

python read_raw.py --dataset_option YOUR_DATASET --metadata_file YOUR_METADATA.csv

If you are using the COPDGene dataset you will find the copd_metadata.csv you need in this repository.

• Run python read_raw.py --dataset_option train, to convert raw train images to nii.gz.
• Run python read_raw.py --dataset_option test, to convert raw test images to nii.gz.

(Remember to set the current directory to the directory where you cloned this repository)

2. Normalization and local contrast adjustment (CLAHE) of all the images.

The preprocessing/preprocessing.py contains the CT_normalization function and the CT_CLAHE function. The CT_normalization function takes the inhale and exhale volumes and does Min-Max Normalization of the whole volume to scale it to the range of (-2000, 2000). Then the values that are greater than -2000 are again scaled but this time to the range of (-1000, 1000). This is made to ensure that the values of the voxels belonging to air are around -1000 (necessary for U-Net lung segmentation). The CT_CLAHE function does contrast limited adaptive histogram equalization on the whole CT volume.

The preprocessing.py script saves the preprocessed images in a directory with the name <dataset_option>_<preprocessing_type>. The preprocessing types available are Normalized, CLAHE, Normalized_CLAHE (does CT_CLAHE on the output of CT_normalization) or CLAHE_Normalized (does CT_normalization on the output of CT_CLAHE).

Calling preprocessing.py from the terminal:

python -m preprocessing.preprocessing  --dataset_option DATASET --preprocessing_type PREPROCESSING_OPTION 

3. Gantry removal

The preprocessing/segment.py script performs gantry removal on the images to keep only the body of the patient. Gantry mask is obtained using k-means clustering followed by post-processing using morphological operations and area filtering of contours. The gantry removed images are saved in data/<dataset_option>_gantry_removed.

Calling segment.py from the terminal:

python -m preprocessing.segment  --dataset_option DATASET --save_gantry_removed True --save_lung_mask False

Additionally, body segmentation masks can be obtained by using the following command:

python -m preprocessing.segment  --dataset_option DATASET --save_gantry_removed False --save_lung_mask False --mask_creation True

4. Segmentation of the lungs

Two methods have been implemented for segmentation of the lungs inside the gantry.

1. K-means based segmentation followed by morphological post-processing.
2. U-Net based segmentation adapted from https://github.com/JoHof/lungmask.

U-Net based segmentation, as expected, performs better. However, for any back-up cases our segmentation method can be used.

Performing lung segmentation with k-means:

 python -m preprocessing.segment --dataset_option DATASET --save_gantry_removed False --save_lung_mask True

Performing lung segmentation with U-Net:

python -m preprocessing.segment_unet --dataset_option DATASET_gantry_removed

Registration

1. Registration of the images (fixed image: inhale, moving image: exhale) using elastix.

By running the function call "elastix_batch_file" located in utils/batchfilecreator.py, a system file is created (.bat or .sh). This elastix system file is ready to perform the registration of the inhale (copdX_iBHCT.nii.gz) to the exhale (copdX_eBHCT.nii.gz) lung images from the desired data/-DATASET by using the desired parameter files in elastix/parameters/PARAMETER_FOLDER and if --mask TRUE using also the provided segmentation in --mask_name either lung_unet, lung_ours or body.

An example of PARAMETER_FOLDER can be found in elastix/parameters/ParOurs. This folder contains two parameters files that elastix uses to perform the registration of the moving exhale image of the lungs into the fixed inhale image. Using to different transformations Affine and Bspline.

python utils/batchfile_creator.py --batch_type elastix --name_experiment NAME_EXPERIMENT
 --parameter PARAMETER_FOLDER --dataset_option -DATASET --mask BOOLEAN --mask_name -MASK_NAME

Once the -DATASET (train) is placed in the working directory as described in the section Structure of the data in our project folder. Run the following line of code to create an example of elastix system file.

python utils/batchfile_creator.py --batch_type elastix --name_experiment example
 --parameter ParOurs --dataset_option train

If the segmentation, for example coming from the U-net, is provided as described in the subsection Segmentation of the lungs in the Data preprocessing. Run the following line of code to create an example of elastix system file.

python utils/batchfile_creator.py --batch_type elastix --name_experiment example_segmentation --parameter ParOurs
--dataset_option train --mask True --mask_name lung_unet

This system file (elastix*.bat or elastix*.sh) is saved in the folder elastix/bat_files.

2. Registration of the landmarks (fixed landmarks: inhale) using transformix

By running the function call "transformix_batch_file" located in utils/batchfilecreator.py, a system file is created (.bat or .sh). This elastix file is ready to perform the transformation of the inhale landmarks applying the last TransformParameters.X.txt that outputs the registration of the images performed by elastix located in elastix/Outputs_experiments_elastix/NAME_EXPERIMENT.

python utils/batchfile_creator.py --batch_type transformix --name_experiment NAME_EXPERIMENT
--parameter PARAMETER_FOLDER --dataset_option -DATASET

Run the following line of code to create an example of elastix system file to run trasnformix.

python utils/batchfile_creator.py --batch_type transformix --name_experiment example
 --parameter ParOurs --dataset_option train

This system file (transformix*.bat or transformix*.sh) is saved in the folder elastix/bat_files. This file outputs in the folder elastix/Outputs_experiments_transformix/NAME_EXPERIMENT different folders containing the transformed inhale landmarks as outputpoints.txt files to be located correctly in the exhale image used in the previous registration with the inhale image.

Compute the metrics

If the exhale landmarks copdX_300_eBH_xyz_r1.txt (groundtruth) files per patient are provided, to check the results coming from the transformation of the inhale landmarks. Running the following line of code will create a .csv file in metrics/ computing the mean TRE and std TRE per patient and the mean and std of all patients between the transformed inhale landmarks and the groundtruth. .

python utils/metrics.py --folder_experiment_landmarks -FOLDER_NAME_OUTPUT_TRANSFORMIX

Our Pipeline

To reproduce the registration of the lung images and its landmarks as we did (results shown in the table below), please follow the command below:

Case ID	Mean TRE	TRE Std Dev
copd1	1.4667	1.6677
copd2	2.5969	2.9965
copd3	1.2824	1.0320
copd4	1.7728	1.3609
mean	1.7797	1.7643

TRAIN DATASET: Lung segmentation

python -m preprocessing.preprocessing  --dataset_option train --preprocessing_type CLAHE_Normalized
python -m preprocessing.segment --dataset_option train_CLAHE_Normalized --save_gantry_removed True --save_lung_mask False
python -m preprocessing.segment_unet --dataset_option train_CLAHE_Normalized_gantry_removed

TRAIN DATASET: Image registration

python -m preprocessing.preprocessing  --dataset_option train --preprocessing_type Normalized_CLAHE
python utils/batchfile_creator.py --batch_type elastix --name_experiment train_Normalized_CLAHE --parameter ParOurs --dataset_option train_Normalized_CLAHE --mask True --mask_name lung_unet
python utils/batchfile_creator.py --batch_type transformix --name_experiment train_Normalized_CLAHE --parameter ParOurs --dataset_option train_Normalized_CLAHE
call elastix/bat_files/elastix_train_Normalized_CLAHE.bat
call elastix/bat_files/transformix_train_Normalized_CLAHE.bat
python utils/metrics.py --folder_experiment_landmarks train_Normalized_CLAHE
python utils/save_outputpoints.py --folder_experiment_landmarks train_Normalized_CLAHE

TEST DATASET: Lung segmentation

python -m preprocessing.preprocessing  --dataset_option test --preprocessing_type CLAHE_Normalized
python -m preprocessing.segment --dataset_option test_CLAHE_Normalized --save_gantry_removed True --save_lung_mask False
python -m preprocessing.segment_unet --dataset_option test_CLAHE_Normalized_gantry_removed

TEST DATASET: Image registration

python -m preprocessing.preprocessing  --dataset_option test --preprocessing_type Normalized_CLAHE
python utils/batchfile_creator.py --batch_type elastix --name_experiment test_Normalized_CLAHE --parameter ParOurs --dataset_option test_Normalized_CLAHE --mask True --mask_name lung_unet
python utils/batchfile_creator.py --batch_type transformix --name_experiment test_Normalized_CLAHE --parameter ParOurs --dataset_option test_Normalized_CLAHE
call elastix/bat_files/elastix_test_Normalized_CLAHE.bat
call elastix/bat_files/transformix_test_Normalized_CLAHE.bat
python utils/save_outputpoints.py --folder_experiment_landmarks test_Normalized_CLAHE