Towards Objective Obstetric Ultrasound Assessment: Contrastive Representation Learning for Fetal Movement Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Prenatal Health

Background:

• Accurate fetal movement (FM) detection is vital for prenatal health assessment.
• Traditional FM detection methods like maternal perception and cardiotocography (CTG) have limitations in accuracy and objectivity.
• Abnormal fetal movement patterns can signal serious complications, including placental dysfunction or fetal distress.

Purpose of the Study:

• To introduce Contrastive Ultrasound Video Representation Learning (CURL), a novel self-supervised learning framework for FM detection.
• To develop a robust method for analyzing fetal movement from extended ultrasound video recordings.
• To improve the objectivity and reliability of FM analysis in prenatal care.

Main Methods:

• CURL utilizes a dual-contrastive loss, combining spatial and temporal contrastive learning for robust motion representation.
• A task-specific sampling strategy is employed for effective separation of movement and non-movement segments during self-supervised training.
• Probabilistic fine-tuning enables flexible inference on arbitrarily long ultrasound recordings.

Main Results:

• CURL achieved a sensitivity of 78.01% and an Area Under the Receiver Operating Characteristic Curve (AUROC) of 81.60% on an in-house dataset.
• The framework demonstrated potential for reliable and objective FM analysis.
• The study highlights the effectiveness of self-supervised contrastive learning for fetal movement analysis.

Conclusions:

• Self-supervised contrastive learning, as implemented in CURL, offers a promising approach for objective fetal movement analysis.
• The developed framework can contribute to improved prenatal monitoring and clinical decision-making.
• Further research in this area could lead to enhanced diagnostic tools for fetal well-being.