Label augmentation method for medical landmark detection in hip radiograph images
Yehyun Suh, Peter Chan, J. Ryan Martin, and Daniel Moyer.
- Converting DICOM format into PNG format: https://github.com/yehyunsuh/DICOM
- Annotating landmarks: https://github.com/yehyunsuh/Landmark-Annotator
- Ubuntu 22.04
- CUDA 11.7
- PyTorch 1.13.0
- Environment Setup
conda create -n label_aug python=3.10 -y
conda activate label_aug
If you do not have conda downloaded in your setup, please refer to conda installation page.
- Clone this repository and set up directories
git clone https://github.com/vine-lab-vu/Label-Augmentation.git
cd Label-Augmentation
mkdir data && cd data
mkdir -p image/all txt && cd ..
- Put your data in the directories
Label-Augmentation
├─ data
│ ├─ image
│ │ ├─ all
│ │ │ ├─ 1.png
│ │ │ ├─ 2.png
│ │ │ ├─ ...
│ │ │ └─ < here goes all the images >
│ └─ txt
│ ├─ test.txt
│ └─ train.txt
├─ utility
│ ├─ dataset.py
│ ├─ log.py
│ ├─ main.py
│ ├─ model.py
│ ├─ preprocess.py
│ ├─ train.py
│ └─ visualization.py
├─ dataset.py
├─ main.py
├─ test.py
└─ train.py
train.txt and test.txt come from Landmark-Annotator
- Download libraries
pip3 install -r requirements.txt
- Start training
python3 main.py --dilate number_of_dilation --dilation_decrease number_of_decrease_in_dilation --dilation_epoch how_many_epochs_per_each_dilation --image_resize size_of_resized_image --batch_size size_of_each_batch --output_channel number_of_labels
If it is your first time training or have added new data, add --preprocess at the end of the command
python3 main.py --test --output_channel number_of_labels
If you have changed any other arguments that is related to the model, you have to add it to the test command.
This repository is built using the segmentation-models-pytorch library.
Yehyun Suh, Peter Chan, J. Ryan Martin, and Daniel Moyer. Label augmentation method for medical landmark detection in hip radiograph images, 2023.
@misc{
suh2023label,
title={Label Augmentation Method for Medical Landmark Detection in Hip Radiograph Images},
author={Yehyun Suh and Peter Chan and J. Ryan Martin and Daniel Moyer},
year={2023},
eprint={2309.16066},
archivePrefix={arXiv},
primaryClass={cs.LG}
}