Improving reliability of three-dimensional fetal echocardiography through motion corrected slice-to-volume registration and machine learning

The prenatal detection of congenital heart disease (CHD) continues to rely on a single imaging modality – 2D ultrasound (2DUS) – for both screening and diagnosis, with a heavy reliance on operator expertise. Advanced 3D ultrasound techniques, which may aid expert diagnosis of complex forms of CHD, are technically limited with poor reliability in clinical practice. This project intends to combine our extensive clinical and biomedical engineering experience to generate a novel acquisition and image-processing pipeline, using deep learning methods to recompile free-hand 2DUS images into isotropic, high spatiotemporal resolution advanced 3D fetal echo (a3DFE) datasets. By leveraging oversampled data acquired in multiple orientations, we will reduce reliance on a single imaging window, improving both reliability and reproducibility. We will then extend our reconstruction network to perform automated segmentation and extraction of functional metrics, coupled with cross-modality validation experiments in preparation for translation into clinical practice. We believe these novel methods will transform expert-level diagnosis of CHD, as well as offering the potential to transform approaches to population-based screening in the future.