This repository contains the official code for the following paper:
@article{madesta:2024,
doi = {XXX},
year = {2024},
month = {XXX},
publisher = {XXX},
volume = {XXX},
number = {XXX},
pages = {XXX},
author = {Frederic Madesta, Thilo Sentker, Clemens Rohling, Tobias Gauer, and Ren\'{e} Werner},
title = {Monte Carlo-based simulation of virtual 3D and 4D Cone-Beam CT from CT images: An end-to-end framework and a novel deep learning-based speedup strategy},
journal = {XXX}
}
This package contains all the required code to perform the following tasks in a fully automated end-to-end fashion:
- Deep learning-based segmentation of CT images into various organ and tissue classes
- Monte Carlo simulation of 3D CBCT images from CT images
- Monte Carlo simulation of 4D CBCT images from CT images with respective correspondence models and respiratory signals
- Reconstruction of 3D and 4D CBCT images from simulated projections using the Reconstruction Toolkit (RTK)
Both the MC as well as the reconstruction code are shipped as pre-compiled binaries in a Docker image for user experience reasons.
The Docker image can be built using the following command:
sh ./build-docker.shThis will build the Docker image with the name cbct-mc and the tag latest.
The Python package can be installed using the following command:
pip install -e .Make sure to execute this command in the root directory of the repository (where the setup.py file is located).
In general, the framework requires very litte data preparation:
- The CT images should be stored in a single-file format (e.g.
*.nii,*.mhaor any other format that can be read by ITK) - The CT images should have non-preporcessed Hounsfield units (HU), i.e. the HU values should be in the range [-1024, 3071]
The MC simulations are performed using the following command (both 3D and 4D simulations):
(4d-cbct-mc) fmadesta@hydrogen:~/research/4d-cbct-mc$ run-mc --help
Usage: run-mc [OPTIONS]
Options:
--image-filepath FILE CT image to use for simulation
--geometry-filepath FILE Geometry to use for simulation. Can be
provided instead of CT image.
--output-folder DIRECTORY Output folder for simulation results
--simulation-name TEXT Name of the simulation. If not provided, the
name is derived from the image filepath.
--gpu INTEGER GPU PCI bus ID to use for simulation (can be
checked with nvidia-smi) [default: 0]
--reference Enable reference simulation
--reference-n-histories INTEGER
Number of histories for reference simulation
[default: 11903320312]
--speedups FLOAT Speedup factors for simulation
--speedup-weights FILE Weights file for speedup model [default:
/home/fmadesta/research/4d-cbct-
mc/cbctmc/assets/models/speedup/default.pth]
--segmenter-weights FILE Weights file for the segmenter model
--segmenter-patch-shape <INTEGER INTEGER INTEGER>...
Patch shape for the segmenter model
[default: 256, 256, 128]
--segmenter-patch-overlap FLOAT RANGE
Overlap ratio for patch-based segmentation
[default: 0.5; 0.0<x<=1.0]
--n-projections INTEGER Number of projections for simulation
[default: 894]
--reconstruct-3d Enable 3D reconstruction
--reconstruct-4d Enable 4D reconstruction
--forward-projection Enable forward projection
--no-clean Disable cleaning of intermediate files
--correspondence-model FILE Correspondence model file. Must be provided
for 4D simulation.
--respiratory-signal FILE Respiratory signal file. Must be provided
for 4D simulation.
--respiratory-signal-quantization INTEGER
Quantization level for respiratory signal. A
lower value means that the respiratory
signal is more coarse.
--respiratory-signal-scaling FLOAT
Scaling factor for respiratory signal
[default: 1.0]
--precompile-geometries Precompile geometries for 4D simulation
--cirs-phantom Use CIRS phantom for simulation
--catphan-phantom Use Catphan604 phantom for simulation
--dry-run Perform a dry run without executing the
simulation
--random-seed INTEGER Random seed for simulation [default: 42]
--loglevel [debug|info|warning|error|critical]
Logging level [default: info]
--help Show this message and exit.
A 3D CBCT simulation is defined by the static (patient) geometry and the moving CBCT scan geometry (i.e. X-ray source and detector).
Analog to the 3D CBCT simulation, a 4D CBCT simulation is defined by the time-resolved/dynamic (patient) geometry and the moving CBCT scan geometry (i.e. X-ray source and detector). In addition, a 4D CBCT simulation requires a correspondence model and a respiratory signal. Thus, the run-mc command has to be called with the --correspondence-model and --respiratory-signal arguments. The correspondence model can be fitted using the fit-correspondence-model command (see below). The respiratory signal can be obtained from a 4D CT scan.
The correspondence model can be fitted using a 4D CT and the corresponding respiratory signal. If no respiratory signal is available, the lung volume can be used as a surrogate signal. The correspondence model is readily fitted by the following code snippet:
import numpy as np
from cbctmc.registration.correspondence import CorrespondenceModel
images: np.ndarray
masks: np.ndarray
timepoints: np.ndarray
model = CorrespondenceModel.build_default(
images=images,
masks=masks,
timepoints=timepoints,
masked_registration=False,
device="cuda:0",
)
model.save("/some/folder/correspondence_model.pkl")