This is the official implementation of our Pattern Recognition journal paper CurvNet: Latent Contour Representation and Iterative Data Engine for Curvature Angle Estimation, which is built upon mmocr-0.2.1. CurvNet is a framework with self-generative capabilities for automatic measurement of scoliosis Cobb angles.

Regardless of your specific needs, you will need to complete the following dependency installations. It is recommended to use a virtual conda environment to manage your environment setup. Run the following commands to install dependencies.

conda create -n CurvNet python=3.7 -y
conda activate CurvNet
conda install pytorch=1.8 torchvision cudatoolkit=11.1 -c pytorch -c nvidia -c conda-forge
pip install mmcv-full==1.3.9 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.8.0/index.html
pip install mmdet==2.14.0
git clone https://github.com/Ernestchenchen/CurvNet
cd CurvNet
pip install -r requirements.txt
python setup.py build develop

You have the following options for your spinal X-ray dataset:

Prepare your own private spinal X-ray dataset. Download our open-source Spinal-AI2024 dataset. Additionally, our code provides sample data that can be used for testing purposes.

Please download and extract the above datasets into the data folder following the file structure below.

data
├─Spinal-AI2024
│  ├─Spinal-AI2024_train.json
│  └─train
│  ├─Spinal-AI2024_test.json
│  └─test
├─Your_Custom_Dataset
│  ├─Your_Custom_train.json
│  └─train
│  ├─Your_Custom_test.json
│  └─test

Ask CurvNet to generate an X-ray image for you (based on the distribution characteristics of our private dataset Spinal2023), and you can continue to complete Function 3, or generate an entire dataset and continue to complete Function 4.

Before executing the instructions, you need to place the images or dataset in the original-X-ray folder, and wait for the output results from the x-ray_generation/x-ray_generate folder.

python x-ray_generation/x-ray_generate/xray_test_noinput.py

Provide a private spine X-ray dataset, train the generation module, and CurvNet to generate X-ray images based on your own data features, continuing to complete Function 3 or Function 4.

You may need to re-adjust the weight file paths or re-initialize the weights to obtain weights that are suitable for your own dataset.

python x-ray_generation/x-ray_generate/train_unet_only_Xray_Pipeline.py

Provide a private or CurvNet generated spine X-ray image, and CurvNet will automatically detect spinal landmarks, Cobb angles, and visualize the results.

CUDA_VISIBLE_DEVICES=0 python tools/test.py configs/spinal_det.py work_dirs/spinal_det/latest.pth --eval hmean-e2e

Provide your private or CurvNet generated unlabeled spine X-ray dataset, and CurvNet will use the Data Engine proposed in the paper to generate relatively usable data annotations for the dataset.

CUDA_VISIBLE_DEVICES=0 tools/train.py configs/spinal_det.py  --work-dir work_dirs/spinal_det

We sincerely thank MMOCR for their excellent works.