Cervix Ultrasound Texture Analysis to Differentiate Between Term and Preterm Birth Pregnancy: A Machine Learning Approach

Abstract

Preterm birth (PTB) is the leading cause of morbidity and mortality in neonates. Currently, the prediction of PTB is based on the identification of a short (less than 25 mm) cervix length (CL) measured by transvaginal ultrasound (TVUS). However, this methodology suffers a low sensitivity (< 50%). Therefore, there is an unmet need for developing better predictors of PTB. Textural analysis of B-mode US images has shown potential in providing quantitative biomarkers for tissue characterization. In this study, we investigated the utility of a texture-based machine-learning method applied to TVUS images to differentiate between term and preterm delivery and identify the potential risk of PTB. Sagittal TVUS images taken at 28 - 32 weeks of gestation were analyzed, and five regions of interest (ROI) were labeled. Morphological transforms (Prewitt, Sobel) and normalization were applied to the images to generate a vast pool of imaging features. To select the best features for building predictive models, Borda ranking was applied. With the selected features, three classifier models were made: logistic regression (LR), random forest (RF), and multilayer perceptron (MLP). At a fixed false positive rate of 10 percent, the MLP model achieved a sensitivity of 67 percent, suggesting that the Borda ranking procedure has a high potential for selecting meaningful features to be used in simple non-linear models, such as the MLP.