2025 Theses Doctoral

Automated uterocervical feature extraction from transvaginal ultrasound images to inform preterm birth prediction and digital twin models

Dagle, Alicia Breanne

Preterm birth (PTB) is the leading cause of perinatal death, affecting approximately 10% of pregnancies, both in the US and abroad. Despite global health campaigns to address PTB, rates have remained stagnant and even on the rise in some countries over the past decade. Currently, transvaginal ultrasound (TVUS) measurement of cervical length (CL) is the soleclinically-accepted quantitative imaging metric for PTB risk, but offers limited predictive value. Many hospital systems incorporate serial CL screening (multiple time points) to identify risk of sPTB, but others have not adopted this practice because there is a lack of scientific consensus on whether this metric is predictive enough to justify the costs. While cervical length is believed to be an important predictor of PTB, it likely does not capture the entire picture of cervicabiomechanical health and remodeling.

Incorporation of additional uterocervical features, describing the shape and size of maternal anatomy, could help improve prediction models of preterm birth, as well as guide computational models of pregnancy. While computational models of cervical biomechanics show promise as risk predictors of PTB, they require precise, time-intensive clinician-provided measurements. Automatic extraction of anatomical features, following AI-enabled ultrasound segmentation, offers a solution to this labeling bottleneck.

This research aims to 1) train AI-based segmentation models to label cervical anatomy from cervical ultrasound images, 2) extract uterocervical geometry measurements from these labeled anatomy masks, 3) examine preliminary links between newly proposed uterocervical geometry features and PTB outcomes, 4) explore how these geometries can be used to automatically create three dimensional geometries for simulations of pregnancy, and 5) create a large clinical dataset of cervical ultrasound images linked to patient health record data and birth outcomes (gestational age at delivery). Because pregnancy is a protected population, databases of this nature are extremely limited in number and access. Ultimately, the newly curated clinical database can be used to verify the results of this research and support future studies of pregnancy.

This research explores a new paradigm to measure cervical structural features in-vivo, which can be used to power 3D digital twin models of pregnancy, quantitative ultrasound-based stiffness measurements of the cervix, and machine-learning based prediction models of PTB. Additionally, this work can be adapted as a teaching-tool for novice sonographers, and used to increase access to cervical length screening for PTB in low-resource settings. Ultimately, this technology may be deployed in the clinic to guide sonographers taking cervical ultrasound images, and provide human-in-the-loop automation of cervical feature measurements, which may better describe cervical structural health in pregnancy, and therefore elucidate the pathways of preterm birth prevention.