Identification of Crown and Rump in First-Trimester Ultrasound Images Using Deep Convolutional Neural Network

Abstract

First-Trimester Ultrasound scans provide invaluable insight into early pregnancies. The scan is used to estimate the gestational age by providing a measurement of the Crown to Rump Length (CRL), it is a crucial scan as it informs obstetric practitioners of the optimal timing for any necessary interventions at the earliest point. Inter-observer variation creates problems for Obstetric Practitioners as any variation in the measurement of the CRL can carry complications to the fetus’ health. Existing machine learning systems to solve this problem are limited; this work details the creation of a machine learning pipeline that implements three Convolutional Neural Networks models (CNNs) to help identify the Crown and Rump regions in First-Trimester Ultrasound Images. The system segments the fetus in the image using a U-Net Model. The segmented image is then subject to an image classification model that implements a pre-trained CNN model, namely, VGG-16. This model is used to classify the segmented images into ‘Good’ and ‘Bad’. Finally, the segmented images are entered into a pre-trained ResNet34 model that identifies the Crown and Rump regions. This can be used by obstetric practitioners to provide an accurate CRL of the fetus and to comment on the actual development of the fetus from the First-trimester Ultrasound images. The system will mitigate issues with the estimation of the gestational age and reduce the inter-observer variations.