OXnet: Deep Omni-supervised Thoracic DiseaseDetection from Chest X-rays

by Luyang Luo, Hao Chen, Yanning Zhou, Huangjing Lin, and Pheng-Ann Heng

Repository for MICCAI 2021 paper OXnet: Deep Omni-supervised Thoracic DiseaseDetection from Chest X-rays

Preliminary: pytorch-retinanet

This code is modified from pytorch-retinanet. Please refer to INSTALL.md for preliminary requirements. Please also make following directories to save the log files and checkpoints:

mkdir logs
mkdir checkpoints

Dataset preparation

Due to security concern, we cannot release the dataset.

To apply OXnet on a custom dataset:

Note that the data shall follow the COCO style;
Put the the json and image file in a same root directory, e.g., root/of/json/and/image/files;
Modify the path to json file (line 62-63) and the path to image files (line 102-103) in retinanet/dataloader.py;
Modify the set_name in train_oxnet.py (line 92, 94, and 96) or train_retinanet.py (line 72, 74, and 76).
NOTE: To training the current version of OXnet, you have to prepare both labeled data and unlabeled data. You can create two COCO-style json files for labeled data and unlabeled data and merge these two in one single json file. Say you have N labeled data, and M unlabeled data, then the beginning N data in the merged file should be labeled ones and the followed M data are the unlabeled ones. Set N and N+M to be num_labeled_data and num_data in the training code. One may see #1 (comment) for reference.

The data tree is as follows:

    ├── Root
          ├── Images
                ├── image_1.png
                ├── image_2.png
                ├── image_n.png
          ├── Json
                ├── train.json
                ├── val.json
                ├── test.json

Training

Before training OXnet, you may need to pre-train a RetinaNet and use the weights to initialize the teacher and the student models in OXnet. In this way, the teacher part of OXnet could generate meaningfule psuedo labels:

python train_retinanet.py --dataset coco --coco_path /root/of/json/and/image/files --experiment_name retinanet  --depth 101

Note that when there are data of different types of annotations, in the json file, labeled data are listed at the beginning and unlabeled data are listed behind. To train OXnet :

python train_oxnet.py --dataset coco --coco_path root/of/json/and/image/files --load_pth path/to/pretrained/model.pt --num_labeled_data 2725 --num_data 13964 --experiment_name oxnet --depth 101

TODO list

Modify dataloader to load three types of data simultaneously: fully-annotated, weakly-annotated, and unlabeled.

Citation

If you find the paper or the code helpful to your own work, please consider cite:

@inproceedings{luo2021oxnet,
      title={OXnet: Omni-supervised Thoracic Disease Detection from Chest X-rays}, 
      author={Luyang Luo and Hao Chen and Yanning Zhou and Huangjing Lin and Pheng-Ann Heng},
      booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
      year={2021}
}

Acknowledgement

The codes are modified from pytorch-retinanet. Part of data augmentation is adapted from STAC, and part of the loss is adapted from GPA-detection.

Questions

Please contact 'lyluo@cse.cuhk.edu.hk'

LLYXC / OXnet