Features/Networks

• Liver Segmentation UNet: Based on https://doi.org/10.1007/978-3-319-24574-4_28.

• The usual FashionMNIST example, for learning purposes.

Design Principles

Each project herein should provide the following:

• Code that passes pylint.

• Unit testing, as appropriate. In all likelihood, testing will cover individual functions, not large training cycles.

• Sufficient logging, including date, time, software (git) version, runtime folder, machine name.

• A main class containing a network that can be run separately in train/test mode.

• Visualisation with TensorBoard.

• Saving of learned network weights at the end of training.

• Loading of pre-train weights, initialising the network ready for inference.

• The ability to be run repeatedly for hyper-parameter tuning via python scripting, not bash.

• The ability to be callable from within a Jupyter Notebook, and thereby amenable to weekly writup’s for supervisions.

• One or more command line programs that are pip-installable, enabling a subsequent user to train and test your algorithm with almost-zero faff.

• Visualisation for debugging purposes, such as printing example image thumbnails etc. should be done in Jupyter notebooks, or in tensorboard, not in the same class as your algorithm.

Optional features could include:

• Small test projects that train quickly to completion won’t need checkpointing, but large ones will.

Usage

Typical instructions for use:

First create a clean python environment, just installing tox:

# Create a clean conda environment
conda create -n myenv python=3.6
conda activate myenv
pip install tox

Then you get the code, and use tox to install all other dependencies:

git clone https://github.com/UCL/scikit-surgerytf
cd scikit-surgerytf
# edit requirements.txt, changing tensorflow to tensorflow-gpu.
# The default is the CPU version just for cross platform testing,
# but for real use, you should swap it to GPU.
# Then run tox to install all dependencies.
tox

Then you can activate the tox created virtualenv and run top-level entry points directly from the root folder:

source .tox/py36/bin/activate
python sksurgeryrgbunet.py --help

Windows users would run:

.tox\py36\Scripts\activate
python sksurgeryrgbunet.py --help

So, for example, to run the sksurgeryrgbunet.py program and train on some data, you would do:

python sksurgeryrgbunet.py -d DATA -w working_dir -s output.hdf5

where DATA is a directory like:

DATA/P1/masks
DATA/P1/images
DATA/P2/masks
DATA/P2/images
.
.
DATA/PN/masks
DATA/PN/images

and P1,P2..PN just represents some patient identifier. Images and masks, though in different folders, must have the same name.

Developing

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

Installing

You can pip install directly from the repository as follows:

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

Contributing

Please see the contributing guidelines.

Useful links

• Source code repository

• Documentation

Licensing and copyright

Copyright 2019 University College London. scikit-surgerytf is released under the Apache Software License 2.0. Please see the license file for details.

Acknowledgements

Supported by Wellcome and EPSRC.