Unlabeled Data-Driven Fetal Landmark Detection in Intrapartum Ultrasound

This repository contains the implementation of our MICCAI 2025 IUGC Challenge winning solution for fetal landmark detection in intrapartum ultrasound and automatic Angle of Progression (AoP) estimation.

Our method is built on a modified TransUNet with a TinyViT backbone, and further improves performance through:

• MAE-assisted knowledge distillation from an ultrasound foundation model
• Semi-supervised learning with pseudo-labeling
• Label perturbation for device-domain adaptation

On the IUGC2025 Challenge test set, our method achieved:

• Mean Radial Error (MRE): 11.6749 px
• Mean Absolute AoP Error: 3.8061°

---

🧩 Overview

The goal of this work is to automatically detect three anatomical landmarks in intrapartum ultrasound images:

• PS1
• PS2
• FH1

These landmarks are then used to compute the Angle of Progression (AoP), an important clinical parameter for assessing fetal head descent during labor.

Our framework consists of two stages:

1. Pretraining: domain-specific representation learning with MAE-assisted knowledge distillation
2. Main training: heatmap-based landmark detection with TransUNet, enhanced by pseudo-labeling and device-domain adaptation

---

⚙️ Usage

1. Pretraining

Run the code in pretrain/ to pretrain the TinyViT backbone with MAE-assisted knowledge distillation.

2. Training

Run the code in train/ to train the landmark detection model based on the pretrained backbone.

3. Evaluation

Use the trained model to predict landmark heatmaps and compute the final AoP from the detected coordinates.

Detailed commands, environment setup, and dataset preparation instructions will be released soon.

---

📚 Citation

If you find this repository useful, please cite our paper:

@inproceedings{ma2026unlabeled,
  title     = {Unlabeled Data-Driven Fetal Landmark Detection in Intrapartum Ultrasound},
  author    = {Ma, Chen and Li, Yunshu and Guo, Bowen and Jiao, Jing and Huang, Yi and Wang, Yuanyuan and Guo, Yi},
  booktitle = {IUGC 2025},
  series    = {Lecture Notes in Computer Science},
  volume    = {16317},
  pages     = {14--23},
  year      = {2026},
  publisher = {Springer Nature Switzerland},
  doi       = {10.1007/978-3-032-11616-1_2}
}

@article{tinyusfm,
  author={Ma, Chen and Jiao, Jing and Liang, Shuyu and Fu, Junhu and Wang, Qin and Li, Zeju and Wang, Yuanyuan and Guo, Yi},
  journal={IEEE Journal of Biomedical and Health Informatics}, 
  title={TinyUSFM: Towards Compact and Efficient Ultrasound Foundation Models}, 
  year={2026},
  pages={1-14},
  doi={10.1109/JBHI.2026.3678309}
}

---

📝 License

This project is released for academic research only.