A Machine Learning Framework for Fully Automatic 3D Fetal Cardiac Ultrasound Evaluation

Fetal cardiac ultrasound (US) is an understudied but important area of medical image analysis. In this work, we identify sources of error and obstacles that may render artificial intelligence (AI) models ineffective in this particular setting. We then present an efficient AI segmentation pipeline for the fetal heart using raw 3D-US volume data with no prior processing. We applied our work on a dataset consisting of 30 3D-US volumes from 26 participants, acquired using 3 different probes on 2 different ultrasound machines. Using an appropriate data enhancement schema, performance of fetal cardiac segmentation improves using state-of-the-art deep learning (DL) methods. We obtained a 19% increase in the Dice Similarity Coefficient (DSC) for convolutional neural networks (CNN). A 16% increase was observed for transformer based networks. The machine learning framework focuses on the data rather than the method, and is able to achieve good performance in spite of the numerous variations in the dataset.