Automatic evaluation of atlantoaxial subluxation in rheumatoid arthritis by a deep learning model

This code is for Okita, Y., Hirano, T., Wang, B. et al. Automatic evaluation of atlantoaxial subluxation in rheumatoid arthritis by a deep learning model. Arthritis Res Ther 25, 181 (2023).

If you have any questions about this code, please send a message to neck_ra_project[at]is.ids.osaka-u.ac.jp.

Usage

Environment

Please first make the environment following the MMpose instruction. After that install the following package:

pip install opencv-python
pip install pandas
pip install matplotlib

Data Preparation

All setting can be modified in file json_generation.py.
A json file need to be pre-made for training from a csv file and image folder.
csv file: refer to annotation.csv.
Image folder: should with images in the direction "data/NECK/images/neck" + str(pic_id) + ".jpg"
We also manually set the cross-validation (see "split" in line 52, we currently set 10 folds.)
"index" in line 53 can set one fold for validation and others for training (default as 0, use first folder).
After setting, generate the json file by running command:

python json_generation.py

Training

For training, we default using model as "hrnet_w32_coco_tiny_256x192". The total epoch can be set in line 262 (default as 40). And run the command:

python train_own.py

During training, we make an evaluation on validation after each epoch with 'PCK', 'NME', "AUC".

Inference

And run the command:

python inference.py

Showing two image of "Pose Estimation" and "Truth". Note that the color of point and connection can be modified in file "configs/base/datasets/custom.py"

Evaluation on Test Set

python eval_own.py

In line 191, change Json file name for the set you want to evaluate.

Draw Loss Plot

python draw_loss.py