In recent years, deep learning has made significant advances in the field of automated medical landmark detection. However, most studies focus on single-center scenarios or domain adaptation. The acquisition of multi-center data presents challenges including privacy concerns and the high costs of data collection and annotation. To address these challenges, we propose a novel framework for domain-generalized medical landmark detection, leveraging single-center data to enhance domain generalization capabilities.

Our approach utilizes public medical segmentation datasets to pre-train a model focused on regions critical for landmark detection. We introduce a novel boundary-aware pre-training strategy along with a mixing loss function to enhance robustness and generalization.

You should download data from Google Drive Folder and unzip them to the corresponding dir to reproduce the results.

• landmark_pretraining/: Contains code for pre-training the model.
  • pretrain.sh: Bash script to run pre-training.
• landmark_detection/: Contains code for the downstream task of landmark detection.
  • train.sh: Bash script to start training.
  • eval.sh: Bash script to run inference.
• runs/: Directory for saving training weights and logs.

To set up the necessary environment for running the provided code, please use the environment.yml file included in the repository. You can create a Conda environment with the required packages by running the following command in your terminal:

conda env create -f environment.yml

The dataset for the detection task can be downloaded from the provided web disk links. Please place the downloaded data in the appropriate directory for further processing.

Navigate to the landmark_pretraining directory and execute the pre-training script by entering the following in your terminal:

cd landmark_pretraining
./pretrain.sh

After pre-training, navigate to the landmark_detection directory to start training and evaluation. Run the following commands:

cd ../landmark_detection
./train.sh
./eval.sh

To cite our work, please use the following format:

@inproceedings{gong2025domain,
  title={Domain Generalized Medical Landmark Detection via Robust Boundary-aware Pre-training},
  author={Gong, Haifan and Yu, Lu and Wan, Xiang and Li, Haofeng},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  pages={1--9},
  year={2025}
}