E-Pix2Vox++ ·

Repository containing source code to implement Efficient Pix2Vox++ for 3D Cardiac Reconstruction from 2D echo views

Cite this work [Springer]

@InProceedings{10.1007/978-3-031-16902-1_9,
author="Stojanovski, David
and Hermida, Uxio
and Muffoletto, Marica
and Lamata, Pablo
and Beqiri, Arian
and Gomez, Alberto",
editor="Aylward, Stephen
and Noble, J. Alison
and Hu, Yipeng
and Lee, Su-Lin
and Baum, Zachary
and Min, Zhe",
title="Efficient Pix2Vox++ for 3D Cardiac Reconstruction from 2D Echo Views",
booktitle="Simplifying Medical Ultrasound",
year="2022",
publisher="Springer International Publishing",
address="Cham",
pages="86--95",
abstract="Accurate geometric quantification of the human heart is a key step in the diagnosis of numerous cardiac diseases, and in the management of cardiac patients. Ultrasound imaging is the primary modality for cardiac imaging, however acquisition requires high operator skill, and its interpretation and analysis is difficult due to artifacts. Reconstructing cardiac anatomy in 3D can enable discovery of new biomarkers and make imaging less dependent on operator expertise, however most ultrasound systems only have 2D imaging capabilities. We propose both a simple alteration to the Pix2Vox++ networks for a sizeable reduction in memory usage and computational complexity, and a pipeline to perform reconstruction of 3D anatomy from 2D standard cardiac views, effectively enabling 3D anatomical reconstruction from limited 2D data. We evaluate our pipeline using synthetically generated data achieving accurate 3D whole-heart reconstructions (peak intersection over union score {\$}{\$}> 0.88{\$}{\$}>0.88) from just two standard anatomical 2D views of the heart. We also show preliminary results using real echo images.",
isbn="978-3-031-16902-1"
}

Datasets

We use the ShapeNet and "Virtual cohort of 1000 synthetic heart meshes from adult human healthy population" datasets in our experiments, which are available below:

ShapeNet rendering images: http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz
1000 synthetic heart meshes: https://zenodo.org/record/4506930#.YuEntdLMJ1M

Installing

How to set up the appropriate environment to run the code:

git clone https://github.com/david-stojanovski/E-Pix2Vox-reconstruction.git
cd E-Pix2Vox-reconstruction/
pip install -r requirements.txt

Get Started (Data Generation)

For the cardiac mesh data set, after it has been downloaded and all files placed in a single folder, the following should be changed:

In data_extraction/convert_ascii2binary.py:

data_folder = r'/path/to/your/vtkdata/'
save_folder = r'/path/to/your/data_save_folder'

In data_extraction/run_plane_extract.py:

data_folder = '/folder/containing/binary/meshes/'
save_folder = r'/folder/to/save/slices/to'

In data_extraction/voxelize_pyntcloud.py:

data_folder = r'/path/to/heart/meshes/'
save_folder = r'/path/to/save/voxel_models/to'

After this you can run:

python convert_ascii2binary.py
python run_plane_extract.py
python voxelize_pyntcloud.py

You should now have all the primary data required to perform 3D reconstruction using the deep learning networks. These folders should now be moved over to the "/e_pivox/datasets" folder.

Get Started (3D Cardiac Reconstruction)

You must at least update the following in the config.py file:

__C.DATASETS.HEARTSEG.IMG_ROOT = '/home/e_pivox/datasets/heart_seg/heart_render'
__C.DATASETS.HEARTSEG.TAXONOMY_FILE_PATH = '/home/e_pivox/datasets/HeartSeg.json'
__C.DATASETS.HEARTSEG.VOXEL_PATH = '/home/e_pivox/datasets/heart_seg/voxel_volumes/%s/%s/model.npy'

__C.DATASETS.SHAPENET.TAXONOMY_FILE_PATH = r'/home/e_pivox/datasets/ShapeNet.json'
__C.DATASETS.SHAPENET.RENDERING_PATH = r'/home/e_pivox/datasets/ShapeNet/ShapeNetRendering/%s/%s/rendering/%02d.png'
__C.DATASETS.SHAPENET.VOXEL_PATH = r'/home/e_pivox/datasets/ShapeNet/ShapeNetVox32/%s/%s/model.binvox'

To train the networks, you can then run:

python runner.py

Licensing

This project is open sourced under MIT license. (See LICENSE for further details)

david-stojanovski / E-Pix2Vox-reconstruction