Decoupled Unbiased Teacher for Source-Free Domain Adaptive Medical Object Detection

📌 This is an official PyTorch implementation of [TNNLS 2023] - Decoupled Unbiased Teacher for Source-Free Domain Adaptive Medical Object Detection.

Decoupled Unbiased Teacher for Source-Free Domain Adaptive Medical Object Detection
Xinyu Liu, Wuyang Li, Yixuan Yuan
The Chinese Univerisity of Hong Kong, City University of Hong Kong

💡 Installation

Please check INSTALL.md for installation instructions.

🔥 Get Started

Step 1: Data Preparation

Option 1: Download data from official data link and convert to detection format.

1. Download datasets from the following sources:

   Source Domain:

   CVC-ClinicDB

   Target Domain:

   Abnormal Symptoms

2. Change the masks to coco style. Please refer to this link or write a script.

Option 2: Use our preprocessed data

1. Download datasets and corresponding coco format annotations in the following links:

   Source Domain:

   CVC-ClinicDB: gDrive

   Target Domain:

   Abnormal Symptoms: gDrive

Change the dataset dir

Change the dataset dir to your downloaded data path in here.

Step 2: Prepare the source-only model

Download the Source-only model from gDrive.

Step 3: Training DUT

# Train the DUT on abnormal symptoms
python tools/train_net_mcd.py --config ./configs/sf/dut_hcmus.yaml OUTPUT_DIR outputs/dut_hcmus

Step 4: Testing the trained model

# Test the trained model
python tools/train_net_mcd.py --config-file configs/sf/dut_hcmus.yaml SOLVER.TEST_ONLY True MODEL.WEIGHT $YOUR .pth WEIGHT$

📝 Citation

If you find this work or codebase is useful for your research, please give it a star and citation. We sincerely appreciate for your acknowledgments.

@article{liu2023decoupled,
  title={Decoupled Unbiased Teacher for Source-Free Domain Adaptive Medical Object Detection},
  author={Liu, Xinyu and Li, Wuyang and Yuan, Yixuan},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2023},
  publisher={IEEE}
}

A relevant research can be found at SMPT, which was our earlier work on source-free polyp detection.

@article{liu2022source,
  title={A source-free domain adaptive polyp detection framework with style diversification flow},
  author={Liu, Xinyu and Yuan, Yixuan},
  journal={IEEE Transactions on Medical Imaging},
  volume={41},
  number={7},
  pages={1897--1908},
  year={2022},
  publisher={IEEE}
}

🤞 Acknowledgement

The code is based on FCOS. Thanks for the excellent framework. For enquiries please contact xinyuliu@link.cuhk.edu.hk.