huaqinghe / joined

JOINED : Prior Guided Multi-task Learning for Joint Optic Disc/Cup Segmentation and Fovea Detection

JOINED

This repository is an official PyTorch implementation of paper:

JOINED: Prior guided multi-task learning for joint optic disc/cup segmentation and fovea detection. In Medical Imaging with Deep Learning MIDL (2022), 2022.

MIDL 2022

Usage

1. Prepare data

Please go to "./Dataset/README.md" for details.

2. Environment

Please prepare an environment with python >= 3.7, and the use the command pip install -r requirements.txt for the dependencies.

3. Train/Test

• Run the train script on GAMMA dataset. The batch size can be reduce to 8 or 4 to save memory (please also decrease the base_lr linearly), and both can reach similar performance. First coarse stage :

cd Coarse
bash train_smp.sh

After Coarse stage, you need to run the eval.sh to get the cropped data and then run the fine stage :

cd Fine
bash train.sh

• Run the test script on GAMMA dataset. First coarse stage :

cd Coarse
bash eval.sh

After Coarse stage, and get the cropped data, run the fine stage :

cd Fine
bash eval.sh

Citations

@inproceedings{he2022joined,
    title={{JOINED}: Prior Guided Multitask Learning for Joint Optic Disc/Cup Segmentation and Fovea Detection},
    author={Huaqing He and Li Lin and Zhiyuan Cai and Xiaoying Tang},
    booktitle={Medical Imaging with Deep Learning},
    year={2022},
    url={https://openreview.net/forum?id=HU6-t9oKvRW}
}

joined's Issues

code？

Why didn't I see the code？-_-

where is the code?

关于标签的处理&问题讨论

x_fovea = round(fovea_location(i, 3));
y_fovea = round(fovea_location(i, 2));
[X_OD, Y_OD] = find(mask == 0);
x_OD = round((max(X_OD) + min(X_OD)) / 2);
y_OD = round((max(Y_OD) + min(Y_OD)) / 2);
这个位置x_fovea ，y_fovea 是颠倒的吗

关于JOINEDTrans

您好～继JOINED之后又看了您JOINEDTrans这篇文章， 摘要中您给的github代码打开后404，请问您这篇文章的代码还会公开吗？

关于数据增强的问题

您好，请问一下如果要对数据进行旋转翻转等增强的话，是不是需要对坐标点进行相应的操作后再生成heatmap和distmap？而不是直接对原始图像的distmap进行旋转翻转