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

TL;DR

This work addresses the need for objective fetal movement analysis from ultrasound by introducing CURL, a self-supervised framework that learns spatiotemporal FM representations from extended video data via dual spatial and temporal contrastive losses. CURL employs task-specific sampling (clean-cut for learning, sliding-window for inference) and a ViT-based spatiotemporal encoder pre-trained with MAE-ST, achieving competitive FM detection performance on a 92-subject dataset (AUROC 81.60%, sensitivity 78.01%). The study demonstrates that self-supervised contrastive learning can produce robust, generalizable FM features suitable for long recordings and remote monitoring, potentially reducing subjectivity and enabling mobile health deployment. Overall, CURL advances objective FM analysis and supports integration into clinical workflows and home-based prenatal care with portable ultrasound devices.

Abstract

Accurate fetal movement (FM) detection is essential for assessing prenatal health, as abnormal movement patterns can indicate underlying complications such as placental dysfunction or fetal distress. Traditional methods, including maternal perception and cardiotocography (CTG), suffer from subjectivity and limited accuracy. To address these challenges, we propose Contrastive Ultrasound Video Representation Learning (CURL), a novel self-supervised learning framework for FM detection from extended fetal ultrasound video recordings. Our approach leverages a dual-contrastive loss, incorporating both spatial and temporal contrastive learning, to learn robust motion representations. Additionally, we introduce a task-specific sampling strategy, ensuring the effective separation of movement and non-movement segments during self-supervised training, while enabling flexible inference on arbitrarily long ultrasound recordings through a probabilistic fine-tuning approach. Evaluated on an in-house dataset of 92 subjects, each with 30-minute ultrasound sessions, CURL achieves a sensitivity of 78.01% and an AUROC of 81.60%, demonstrating its potential for reliable and objective FM analysis. These results highlight the potential of self-supervised contrastive learning for fetal movement analysis, paving the way for improved prenatal monitoring and clinical decision-making.

Figures (9)

• Figure 1: Overview of the clinical trial workflow and dataset characteristics. Left panel (a) illustrates the data acquisition process, where trial data are collected, annotated and systematically stored in a centralized repository. Right panels (b)–(d) detail the movement features: (b) categorizes fetal movements as annotated by clinical experts, grouping them into super-categories based on motion type (e.g., breathing, panting, and hiccups are classified under respiratory movements); (c) presents the distribution of movement sub-classes on a logarithmic scale; and (d) displays the average duration (in seconds) for each movement sub-type. Panels (e)–(h) summarize the demographic data from clinical trial participants, with (e) showing gestational age (weeks), (f) depicting maternal age (years), (g) indicating maternal weight (kg), and (h) representing maternal height (cm).
• Figure 2: Schematic overview of the proposed CURL framework pipeline, comprising six sequential stages. (1) Video clip extraction is performed using either a clean-cut or sliding-window sampling strategy. (2) Spatio-temporal augmentations are applied to construct paired video clips $(x_i, x_j)$, which are then passed through a shared video encoder $f(\cdot)$ to obtain corresponding feature representations $(h_i, h_j)$. These representations are subsequently routed into two distinct branches. (3) In the spatial contrastive learning branch, an MLP projection head $g_{sc}(\cdot)$ projects the features into a latent space for instance-level spatial contrastive learning. (4) In parallel, the temporal contrastive learning branch uses another MLP projection head $g_{tc}(\cdot)$ to map the features into a space that facilitates the clustering of temporally consistent motion patterns. (5) A fine-tuning phase follows, wherein the framework is adapted for the downstream task via a task-specific head. (6) Finally, during inference, the model processes input clips in a sliding-window fashion using a single forward pass.
• Figure 3: Comparative visualization of key characteristics of clean-cut sampling versus sliding-window sampling. (a)Clean-cut sampling selects distinct, non-overlapping video segments, thereby providing independent samples that capture broader temporal contexts.(b) Sliding-window sampling continuously moves a fixed-size window over the video, generating overlapping segments that offer finer-grained temporal dynamics at the expense of some redundancy.
• Figure 4: Example video clips demonstrating spatial and temporal augmentation techniques. Panels (a)–(c) illustrate spatial augmentations applied to video frames: (a) brightness adjustment, (b) rotation, and (c) median blur filtering. Panels (d)–(f) display temporal augmentations: (d) tube masking, (e) frame masking, and (f) random masking.
• Figure 5: (a) Bar graph comparing the performance on fetal movement subclasses among state-of-the-art representation learning algorithms, (b) Box plots presenting the performance distributions of various representation learning algorithms evaluated using 5-fold cross-validation on the fetal movement dataset and (c) ROC curve comparing the proposed algorithm with other state-of-the-art representation learning methods on our in-house dataset, demonstrating superior classification performance. The model used is ViT-L.