The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset

Introduction

This repo contains the implementation of the deep learning-based knee MRI cartilage segmentation model used for osteoarthritis research as described in our paper: The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset. By using this implementation, you can either train new knee MRI cartilage segmentation models or obtain knee MRI cartilage segmentation by using our pretrained model.

Instructions

1. Please refer to IBM Power machine's "PowerAI 1.6.0" conda environment with the dependencies provided in requirements.txt. For Linux machines, please use the requirements.yml file to create a Conda environment:

conda env create -f requirements.yml

2. ./data folder contains subfolders named ./train, ./test , and ./valid which needs to include train, test and validation set images and segmentation maps, respectively. To request the dataset used for the IWOAI challenge, please contact Dr. Chaudhari.
3. Once you have data ready, you can use train.py to train a segmentation model with predefined settings that were used on model training for submission to the IWOAI challenge. Trained models will be saved within the folder named ./TrainedModels
4. Use inference.py file to obtain knee MRI cartilage segmentation by using our pretrained model. MR images located inside ./data/test folder will be used to produce cartilage segmentation maps.

Repo Structure

• ./data: Folder for knee MR images and cartilage segmentation that will be used for training.
• ./InferenceModel: Trained model weights used in this study. They can be downloaded from here.
• ./InferenceResults: will contain the cartilage segmentation maps when inference.py file runs.
• ./TrainedModels: will contain trained cartilage segmentation models when train.py file runs.

Training a DL model

You can use this repo to train models for segmenting the knee MRI cartilage. Default parameters are defined within train.py file. So, you directly run the following script to train a DL model.

python train.py

Please see the arguments in the .py file to make input argument changes if you need.

Inference

Once you download pretrained model weights from here to the ./InferenceModel folder, you can use the following script to obtain cartilage segmentation of knee MR images located under ./data/test:

python inference.py 

Computer System

We used an HPC system with multiple IBM Witherspoon GPU nodes, each node has multiple IBM Power9 4Ghz Processors and V100 NVLink (Volta) GPU’s, and a 512GB RAM. In this project, we used TensorFlow v1.13.1 with a single Nvidia Tesla V100 32GB RAM GPU.

License

This repository is licensed under the terms of the GNU AGPLv3 license.

Reference

If you found this code useful, please cite our paper (our team is highlighted in bold text):

The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset Arjun D. Desai, Francesco Caliva, Claudia Iriondo, Naji Khosravan, Aliasghar Mortazi, Sachin Jambawalikar, Drew Torigian, Jutta Ellermann, Mehmet Akcakaya, Ulas Bagci, Radhika Tibrewala, Io Flament, Matthew O' Brien, Sharmila Majumdar, Mathias Perslev, Akshay Pai, Christian Igel, Erik B. Dam, Sibaji Gaj, Mingrui Yang, Kunio Nakamura, Xiaojuan Li, Cem M. Deniz, Vladimir Juras, Ravinder Regatte, Garry E. Gold, Brian A. Hargreaves, Valentina Pedoia, Akshay S. Chaudhari Radiology: Artificial Intelligence 2021 3:3

@article{desai2021IWOAI,
    title = {The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset},
    author = {Arjun D. Desai, Francesco Caliva, Claudia Iriondo, Naji Khosravan, Aliasghar Mortazi, Sachin Jambawalikar, Drew Torigian, Jutta Ellermann, Mehmet Akcakaya, Ulas Bagci, Radhika Tibrewala, Io Flament, Matthew O' Brien, Sharmila Majumdar, Mathias Perslev, Akshay Pai, Christian Igel, Erik B. Dam, Sibaji Gaj, Mingrui Yang, Kunio Nakamura, Xiaojuan Li, **Cem M. Deniz**, Vladimir Juras, Ravinder Regatte, Garry E. Gold, Brian A. Hargreaves, Valentina Pedoia, Akshay S. Chaudhari}, 
    journal = {Radiology: Artificial Intelligence},
    year = {2021},
    volume = {3},
    number = {3}
    URL = {https://doi.org/10.1148/ryai.2021200078}
}