Generating cervical anatomy labels using a deep ensemble multi-class segmentation model applied to transvaginal ultrasound images

Preterm birth (PTB) is the leading cause of perinatal death, affecting 10% of pregnancies. Currently, transvaginal ultrasound (TVUS) measurement of cervical length (CL) is the sole quantitative imaging metric for PTB risk, but offers limited predictive value. While computational models of cervical biomechanics show promise as PTB risk predictors, they require precise clinician-provided measurements. AI-enabled ultrasound segmentation offers a solution by automatically extracting anatomical features, thus addressing the labeling bottleneck. This study utilizes an ensemble of deep learning-based multi-class segmentation models trained on diverse TVUS data (N = 246) and evaluated on an out-of-distribution dataset (N = 29). High agreement (Dice metric ~ 0.8) between expert and model labels demonstrates the utility of AI tools in accurately measuring cervical geometry. Ultimately, this can enhance biomechanical models and more sophisticated AI-based models to better predict birth timing, specifically targeting PTB risk.