Summary: Researchers from Columbia University are finding ways to develop a computational tool developed to better segment the entire cross-sectional area of the cervix tissue from 2D transvaginal ultrasounds, enabling the creation of a generalizable, cervical-features-based prediction model of PTB risk.
Challenge: A safe, full-term pregnancy is vital to the health and wellbeing of every child, yet it is far from guaranteed. Preterm birth (PTB) is the leading cause of perinatal death and remains a major global health concern. Due to limited pregnancy-related research, clinicians cannot fully explain what triggers healthy, gestationally-appropriate labor, let alone risky premature labor. This lack of fundamental understanding hinders the ability to predict PTB on an individual patient level. This work focuses on the cervix, a complex biomechanical barrier in pregnancy. Sonographic measurement of cervical length with transvaginal ultrasound (TVUS) is a common clinical test to assess the risk of subsequent PTB.
Manual segmentation methods are usually time consuming, not scaleable, and can vary between clinicians and sonographers. Therefore, the researchers chose to adopt an automatic deep learning based multi-class residual UNet architecture segmentation method.
Findings: The researchers demonstrated that the standard-of-care TVUS may be used to accurately segment cervical geometry, enabling the study of cervical variations across pregnancies with broader implications in understanding and ultimately preventing PTB. Given that accurate sonographic cervix segmentation followed by multi-dimensional cervical geometric feature extraction is can have a higher PTB predictive capability compared to sonographic cervical length alone.
How Labelbox was used: All data was annotated using Labelbox, with the model being trained on data from the Cervical Length Education and Review (CLEAR) program, which includes second and third trimester TVUS images from multiple sites and ultrasound systems across the United States.
Read the full paper here.