JS2

Implementation of "A Lightweight CNN and Joint Shape-Joint Space (JS2) Descriptor for Radiological Osteoarthritis Detection", MIUA 2020 paper.

https://arxiv.org/abs/2005.11715

Data:

Multicenter Osteoarthritis Study (MOST): http://most.ucsf.edu/

The Osteoarthritis Initiative (OAI): https://nda.nih.gov/oai/

Setup

1. Obtain OAI and MOST Data from the links above. We can not distribute the data.

2. Extract Landmarks using BoneFinder(BF) software: http://bone-finder.com/.

3. Localize both left and right knee joints using BF landmarks.

4. Save each joint and associated list of landmark points in a seperate .npy file (patientID_SIDE.npy, i.e. "9000099_L.npy").

5. Create a conda environment as follows:

   conda env create -f env.yml

6. Prepare medial tibia crops and JSW measurements using prepare_crops.py. This will save marginal medial tibia ROI and JSW measurements (minJSW, fJSW, JS2, jointsize) as .npy files.

Run

Select model according to models defined in models.py

python main.py --model combined

License

Please cite:

@article{bayramoglu2020js2,
  title={A Lightweight CNN and Joint Shape-Joint Space (JS2) Descriptor for Radiological Osteoarthritis Detection},
  author={Bayramoglu, Neslihan and Miika, Nieminen and Saarakkala, Simo},
  journal={MIUA 2020, arXiv preprint },
  year={2020}
}```