Learning Shape Representation on Sparse Point Clouds for Volumetric Image Segmentation

This repository contains code for the MICCAI 2019 paper "Learning Shape Representation on Sparse Point Clouds for Volumetric Image Segmentation", which can be found at https://doi.org/10.1007/978-3-030-32245-8_31.

Installation

The installation has been tested with Ubuntu 16.04, Python 3.6, TensorFlow 1.10, and CUDA 9.0. The setup.py file lists all other dependencies.

First, create a virtual environment named pc with Python 3.6:

    $ virtualenv --python=python3.6 pc
    $ source ./pc/bin/activate

Second, copy the code:

    $ git clone https://github.com/fabianbalsiger/point-cloud-segmentation-miccai2019
    $ cd point-cloud-segmentation-miccai219

Third, install the required libraries:

    $ pip install -r requirements.txt

Fourth, compile the required TensorFlow extension for farthest point sampling. Note that you might need to edit compile.sh such that the compilation works.

    $ cd pc/model/sampling
    $ ./compile.sh

This should install all required dependencies. Please refer to the official TensorFlow documentation on how to use TensorFlow with GPU support.

Usage

We shortly describe the training procedure. The data used in the paper is not publicly available. But, we provide a script to generate dummy data such that you are able to run the code.

Dummy Data Generation and Configuration

We handle the data using pymia. Therefore, we need to create a hierarchical data format (HDF) file to have easy and fast access to our data during the training and testing. Create the dummy data by

    $ python ./snippets/create_dataset.py

This will create the file ./data/data.h5, or simply our dataset. Use any open source HDF viewer to inspect the file (e.g., HDFView). Please refer to the pymia documentation on how to create your own dataset.

Now, we create a training/validation/testing split file by

    $ python ./snippets/create_split.py

This will create the file ./data/split1_04-02-02.json.

Finally, you need to adjust the paths in the configuration file ./bin/config.json:

• database_file
• result_dir
• model_dir
• split_file

Training

To train the model, simply execute ./bin/main.py. The data and training parameters are provided by the ./bin/config.json, which you can adapt to your needs. Note that you might want to specify the CUDA device by

    $ CUDA_VISIBLE_DEVICES=0 python ./bin/main.py

The script will automatically use the training subjects defined in ./data/split1_04-02-02.json and evaluate the model's performance after each epoch on the validation subjects. The validation will be saved under the path result_dir specified in the configuration file ./bin/config.json. The trained model will be saved under the path model_dir specified in the configuration file ./bin/config.json. Further, the script logs the training and validation progress for visualization using TensorBoard. Start the TensorBoard to observe the training:

    $ tensorboard --logdir=<path to the model_dir>

Support

We leave an explanation of the code as exercise ;-). But if you found a bug or have a specific question, please open an issue or a pull request.

Citation

If you use this work, please cite

Balsiger, F., Soom, Y., Scheidegger, O., & Reyes, M. (2019). Learning Shape Representation on Sparse Point Clouds for Volumetric Image Segmentation. In D. Shen, T. Liu, T. M. Peters, L. H. Staib, C. Essert, S. Zhou, … A. Khan (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2019 (pp. 273–281). https://doi.org/10.1007/978-3-030-32245-8_31

@inproceedings{Balsiger2019b,
address = {Cham},
author = {Balsiger, Fabian and Soom, Yannick and Scheidegger, Olivier and Reyes, Mauricio},
booktitle = {Medical Image Computing and Computer Assisted Intervention – MICCAI 2019},
doi = {10.1007/978-3-030-32245-8_31},
editor = {Shen, Dinggang and Liu, Tianming and Peters, Terry M. and Staib, Lawrence H. and Essert, Caroline and Zhou, Sean and Yap, Pew-Thian and Khan, Ali},
pages = {273--281},
publisher = {Springer},
series = {Lecture Notes in Computer Science},
title = {{Learning Shape Representation on Sparse Point Clouds for Volumetric Image Segmentation}},
volume = {11765},
year = {2019}
}

License

The code is published under the MIT License.