Robust Fetal Pose Estimation across Gestational Ages via Cross-Population Augmentation

Area of Science:

• Medical Imaging
• Fetal Development
• Machine Learning

Background:

• Fetal motion is vital for assessing neurological development and intrauterine health.
• Quantifying fetal motion is challenging, especially in early gestational ages (GA).
• Current pose estimation methods struggle to generalize to early GA due to anatomical changes and data limitations.

Purpose of the Study:

• To develop a data augmentation framework for robust fetal pose estimation across different gestational ages.
• To enable machine learning models to generalize to early GA using data from older GA cohorts.
• To improve the reliability of fetal pose estimation for earlier clinical detection and intervention.

Main Methods:

• Developed a cross-population data augmentation framework for fetal pose estimation.
• Implemented a fetal-specific augmentation strategy simulating early GA intrauterine environments.
• Trained and evaluated pose estimation models using the proposed augmentation on diverse GA cohorts.

Main Results:

• The cross-population augmentation framework significantly improved model generalization to early GA.
• Reduced variability in pose estimation across both older and younger GA cases.
• Demonstrated enhanced robustness of fetal pose estimation throughout gestation.

Conclusions:

• Cross-population data augmentation is effective for improving fetal pose estimation across gestational ages.
• The proposed method facilitates more reliable quantitative analysis of fetal motion.
• This work supports earlier clinical assessment and intervention in high-risk pregnancies using 4D fetal imaging.