scikit-surgerytorch documentation¶

The aim of scikit-surgery torch is to provide a home for various pytorch models/examples/utilities that may be useful for Image Guided Surgery.

Features¶

Implemented models:

High Resolution Stereo network Inference only, see author’s repo for pre trained weights. As at commit aae0b9b.
Volume2SurfaceCNN Inferencece only, see author’s repo for pre trained weights. As at commit 5a656381.
Models can run on GPU or CPU.
Example usage in tests/.

Source code is avaialble on GitHub.

scikit-surgerytorch¶

Author: Thomas Dowrick

scikit-surgerytorch is part of the scikit-surgery software project, developed at the Wellcome EPSRC Centre for Interventional and Surgical Sciences, part of University College London (UCL).

The aim of scikit-surgery torch is to provide a home for various pytorch models/examples/utilities that may be useful for Image Guided Surgery.

Features¶

Implemented models:

High Resolution Stereo network Inference only, see author’s repo for pre trained weights. As at commit aae0b9b.
Volume2SurfaceCNN Inferencece only, see author’s repo for pre trained weights. As at commit 5a656381.
Models can run on GPU or CPU.
Example usage in tests/.

scikit-surgerytorch is NOT meant to be a layer on-top of pytorch or provide a new kind-of platform. The aim is that researchers can learn from examples, and importantly, learn how to deliver an algorithm that can be used by other people out of the box, with just a `pip install`, rather than a new user having to re-implement stuff, or struggle to get someone else’s code running.

Cloning¶

You can clone the repository using the following command:

git clone https://github.com/UCL/scikit-surgerytorch

Running tests¶

Pytest is used for running unit tests:

pip install pytest
python -m pytest

Linting¶

This code conforms to the PEP8 standard. Pylint can be used to analyse the code:

pip install pylint
pylint --rcfile=tests/pylintrc sksurgerytorch

Installing¶

You can pip install directly from the repository as follows:

pip install git+https://github.com/UCL/scikit-surgerytorch

Contributing¶

Please see the contributing guidelines.

Useful links¶

Licensing and copyright¶

Acknowledgements¶

Supported by Wellcome and EPSRC.

Stereo Reconstruction¶

High Resolution Stereo¶

Module to implement Hierarchical Deep Stereo Matching on High Resolution Images network.

class sksurgerytorch.models.high_res_stereo.HSMNet(max_disp: int = 255, entropy_threshold: float = -1, level: int = 1, scale_factor: float = 0.5, weights=None)[source]¶

Class to encapsulate network form ‘Hierarchical Deep Stereo Matching on High Resolution Images’.

Thanks to Gengshang Yang, for their network implementation.

Parameters:

max_disp – Maximum number of disparity levels
entropy_threshold – Pixels with entropy above this value will be ignored in the disparity map. Disabled if set to -1.
level – Set to 1, 2 or 3 to trade off quality of depth estimation against runtime. 1 = best depth estimation, longer runtime, 3 = worst depth estimation, fastest runtime.
scale_factor – Images can be resized before passing to the network, for perfomance impromvents. This sets the scale factor.
weights – Path to trained model weights (.tar file)

predict(left_image: numpy.ndarray, right_image: numpy.ndarray) → numpy.ndarray[source]¶

Predict disparity from a pair of stereo images.

Parameters:	left_image (np.ndarray) – Left stereo image, 3 channel RGB right_image (np.ndarray) – Right stero image, 3 channel RGB
Returns:	Predicted disparity, grayscale
Return type:	np.ndarray

sksurgerytorch.models.high_res_stereo.run_hsmnet_model(max_disp, entropy_threshold, level, scale_factor, weights, left_image, right_image, output_file)[source]¶: This is for the command line entry point

class sksurgerytorch.models.high_res_stereo.toTensorLegacy[source]¶: .

Non Rigid Registration¶

Volume 2 Surface CNN¶

V2SNet Model Impementation

class sksurgerytorch.models.volume_to_surface.Volume2SurfaceCNN(mask: bool = True, weights: str = None, grid_size: int = 64)[source]¶

Class to encapsulate network form ‘Non-Rigid Volume to Surface Registration using a Data-Driven Biomechanical Model’.

Thanks to Micha Pfieffer, for their network implementation.

Parameters:	mask (bool) – If true, use masking weights (str) – Path to trained model weights (.tar file)

predict(preoperative: numpy.ndarray, intraoperative: numpy.ndarray) → numpy.ndarray[source]¶

Predict the displacement field between model and surface.

Parameters:	preoperative (np.ndarray) – Preoperative surface/point cloud intraoperative (np.ndarray) – Intraoperative surface/point cloud
Returns:	Displacement field
Return type:	np.ndarray