Classification of First Trimester Ultrasound Images Using Deep Convolutional Neural Network

Abstract

AbstractFetal ultrasound imaging is commonly used in correctly identifying fetal anatomical structures. This is particularly important in the first-trimester to diagnose any possible fetal malformations. However, inter-observer variation in identifying the correct image can lead to misdiagnosis of fetal growth and hence to aid the sonographers machine learning techniques, such as deep learning, have been increasingly used. This work describes the use of ResNet50, a pretrained deep convolutional neural network model, in classifying \(11-13^{+6}\) weeks Crown to Rump Length (CRL) fetal ultrasound images into correct and incorrect categories. The presented model adopted a skip connection approach to create a deeper network with hyperparameters which were tuned for the task. This article discusses how to distinguish Crown to Rump Length (CRL) fetal ultrasound images into correct and incorrect categories using ResNet50. The presented model used a skip link approach to construct a deeper network with task-specific hyperparameters. The model was applied to a real data set of 900 CRL images, 450 of which were right and 450 of which were incorrect, and it was able to identify the images with an accuracy of 87% on the preparation, validation, and test data sets. This model can be used by the sonographers to identify correct images for CRL measurements and hence help avoid incorrect dating of pregnancies by reducing the inter-observer variation. This can also be used to train sonographers in performing first-trimester scans.KeywordsConvolutional neural networkMachine learningFetal ultrasound imagingCrown to rump lengthDeep learningMedical imaging