Skip to main content

Fast Implementation of Generalised Geodesic Distance Transform for CPU (OpenMP) and GPU (CUDA)

Project description

FastGeodis: Fast Generalised Geodesic Distance Transform

License status CI Build PyPI version

This repository provides CPU (OpenMP) and GPU (CUDA) implementations of Generalised Geodesic Distance Transform in PyTorch for 2D and 3D input data based on parallelisable raster scan ideas from [1]. It includes methods for computing Geodesic, Euclidean distance transform and mixture of both.

2D images, 1 of 4 passes 3D volumes, 1 of 6 passes

The above raster scan method can be parallelised for each row/plane on an available device (CPU or GPU). This leads to significant speed up as compared to existing non-parallelised raster scan implementations (e.g. https://github.com/taigw/GeodisTK). Python interface is provided (using PyTorch) for enabling its use in deep learning and image processing pipelines.

In addition, implementation of generalised version of Geodesic distance transforms along with Geodesic Symmetric Filtering (GSF) is provided for use in interactive segmentation methods, that were originally proposed in [1, 2, 5].

The raster scan based implementation provides a balance towards speed rather than accuracy of Geodesic distance transform and hence results in efficient hardware utilisation. On the other hand, in case of Euclidean distance transform, exact results can be achieved with other packages (albeit not on necessarilly on GPU) [6, 7, 8]

Citation

If you use this code in your research, then please consider citing:

status

Asad, Muhammad, Reuben Dorent, and Tom Vercauteren. "FastGeodis: Fast Generalised Geodesic Distance Transform." Journal of Open Source Software (JOSS), 2022. (paper link)

Bibtex:

@article{asad2022fastgeodis, 
  doi = {10.21105/joss.04532}, 
  url = {https://doi.org/10.21105/joss.04532}, 
  year = {2022}, 
  publisher = {The Open Journal}, 
  volume = {7}, 
  number = {79}, 
  pages = {4532}, 
  author = {Muhammad Asad and Reuben Dorent and Tom Vercauteren}, 
  title = {FastGeodis: Fast Generalised Geodesic Distance Transform}, 
  journal = {Journal of Open Source Software} 
}

Installation instructions

The provided package can be installed using:

pip install FastGeodis

or

pip install git+https://github.com/masadcv/FastGeodis

or (on conda environments with existing installation of PyTorch with CUDA)

pip install FastGeodis --no-build-isolation

Included methods

Optimised Fast Implementations for GPU/CPU based on [1]

Method Description Documentation
Fast Generalised Geodesic Distance 2D Paralellised generalised geodesic distance transform for CPU/GPU [1] FastGeodis.generalised_geodesic2d
Fast Generalised Geodesic Distance 3D Paralellised generalised geodesic distance transform for CPU/GPU [1] FastGeodis.generalised_geodesic3d
Fast Signed Generalised Geodesic Distance 2D Paralellised signed generalised geodesic distance transform for CPU/GPU [1] FastGeodis.signed_generalised_geodesic2d
Fast Signed Generalised Geodesic Distance 3D Paralellised signed generalised geodesic distance transform for CPU/GPU [1] FastGeodis.signed_generalised_geodesic3d
Fast Geodesic Symmetric Filtering 2D Paralellised geodesic symmetric filtering for CPU/GPU [2] FastGeodis.GSF2d
Fast Geodesic Symmetric Filtering 3D Paralellised geodesic symmetric filtering for CPU/GPU [2] FastGeodis.GSF3d

Toivanen's Implementations for CPU based on [9]

Method Description Documentation
Toivanen's Generalised Geodesic Distance 2D Toivanen's generalised geodesic distance transform for CPU [9] FastGeodis.generalised_geodesic2d_toivanen
Toivanen's Generalised Geodesic Distance 3D Toivanen's generalised geodesic distance transform for CPU [9] FastGeodis.generalised_geodesic3d_toivanen
Toivanen's Signed Generalised Geodesic Distance 2D Toivanen's signed generalised geodesic distance transform for CPU [9] FastGeodis.signed_generalised_geodesic2d_toivanen
Toivanen's Signed Generalised Geodesic Distance 3D Toivanen's signed generalised geodesic distance transform for CPU [9] FastGeodis.signed_generalised_geodesic3d_toivanen
Toivanen's Geodesic Symmetric Filtering 2D Toivanen's geodesic symmetric filtering for CPU [2, 9] FastGeodis.GSF2d_toivanen
Toivanen's Geodesic Symmetric Filtering 3D Toivanen's geodesic symmetric filtering for CPU [2, 9] FastGeodis.GSF3d_toivanen

Fast Marching Implementations for CPU based on [4, 10]

Method Description Documentation
Fast Marching Generalised Geodesic Distance 2D Fast Marching generalised geodesic distance transform for CPU [9] FastGeodis.generalised_geodesic2d_fastmarch
Fast Marching Generalised Geodesic Distance 3D Fast Marching generalised geodesic distance transform for CPU [9] FastGeodis.generalised_geodesic3d_fastmarch
Fast Marching Signed Generalised Geodesic Distance 2D Fast Marching signed generalised geodesic distance transform for CPU [9] FastGeodis.signed_generalised_geodesic2d_fastmarch
Fast Marching Signed Generalised Geodesic Distance 3D Fast Marching signed generalised geodesic distance transform for CPU [9] FastGeodis.signed_generalised_geodesic3d_fastmarch
Fast Marching Geodesic Symmetric Filtering 2D Fast Marching geodesic symmetric filtering for CPU [2, 9] FastGeodis.GSF2d_fastmarch
Fast Marching Geodesic Symmetric Filtering 3D Fast Marching geodesic symmetric filtering for CPU [2, 9] FastGeodis.GSF3d_fastmarch

Example usage

Fast Geodesic Distance Transform

The following demonstrates a simple example showing FastGeodis usage:

To compute Geodesic Distance Transform:

device = "cuda" if torch.cuda.is_available() else "cpu"
image = np.asarray(Image.open("data/img2d.png"), np.float32)

image_pt = torch.from_numpy(image).unsqueeze_(0).unsqueeze_(0)
image_pt = image_pt.to(device)
mask_pt = torch.ones_like(image_pt)
mask_pt[..., 100, 100] = 0

v = 1e10
# lamb = 0.0 (Euclidean) or 1.0 (Geodesic) or (0.0, 1.0) (mixture)
lamb = 1.0
iterations = 2
geodesic_dist = FastGeodis.generalised_geodesic2d(
    image_pt, mask_pt, v, lamb, iterations
)
geodesic_dist = np.squeeze(geodesic_dist.cpu().numpy())

To compute Euclidean Distance Transform:

device = "cuda" if torch.cuda.is_available() else "cpu"
image = np.asarray(Image.open("data/img2d.png"), np.float32)

image_pt = torch.from_numpy(image).unsqueeze_(0).unsqueeze_(0)
image_pt = image_pt.to(device)
mask_pt = torch.ones_like(image_pt)
mask_pt[..., 100, 100] = 0

v = 1e10
# lamb = 0.0 (Euclidean) or 1.0 (Geodesic) or (0.0, 1.0) (mixture)
lamb = 0.0
iterations = 2
euclidean_dist = FastGeodis.generalised_geodesic2d(
    image_pt, mask_pt, v, lamb, iterations
)
euclidean_dist = np.squeeze(euclidean_dist.cpu().numpy())

For more usage examples see:

Description Python Colab link
Simple 2D Geodesic and Euclidean Distance samples/simpledemo2d.py Open in Colab
Simple Signed 2D Geodesic and Euclidean Distance samples/simpledemo2d_signed.py Open in Colab
Simple 3D Geodesic and Euclidean Distance samples/simpledemo3d.py Open in Colab
Simple Signed 3D Geodesic and Euclidean Distance samples/simpledemo3d_signed.py Open in Colab
2D Geodesic Distance samples/demo2d.py Open in Colab
2D Signed Geodesic Distance samples/demo2d_signed.py Open in Colab
3D Geodesic Distance samples/demo3d.py Open in Colab
3D Signed Geodesic Distance samples/demo3d_signed.py Open in Colab
2D GSF Segmentation Smoothing samples/demoGSF2d_SmoothingSegExample.ipynb Open in Colab

Unit Tests

A number of unittests are provided, which can be run as:

pip install -r requirements-dev.txt
python -m unittest

Documentation

Further details of each function implemented in FastGeodis can be accessed at the documentation hosted at: https://fastgeodis.readthedocs.io.

Contributing to FastGeodis

Spotted a bug or have a feature request to improve the package? We would love to have your input! See our guidelines for contributing.

Comparison of Execution Time and Accuracy

FastGeodis (CPU/GPU) is compared with existing GeodisTK (https://github.com/taigw/GeodisTK) in terms of execution speed as well as accuracy.

Execution Time

2D images 3D volumes

Accuracy

2D case

Qualitative Comparison Quantitative (joint histogram)

3D case

Qualitative Comparison Quantitative (joint histogram)

References

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

FastGeodis-1.0.2.tar.gz (23.2 kB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page