4D foetal cardiac ultrasound image detection based on deep learning with weakly supervised localisation for rapid diagnosis of evolving hypoplastic left heart syndrome

Abstract

Hypoplastic left heart syndrome (HLHS) is a rare, complex, and incredibly foetal congenital heart disease. To decrease neonatal mortality, evolving HLHS (eHLHS) in pregnant women should be critically diagnosed as soon as possible. However, diagnosis is currently heavily dependent on skilled medical professionals using foetal cardiac ultrasound images, making it difficult to rapidly and easily examine for this disease. Herein, the authors propose a cost-effective deep learning framework for rapid diagnosis of eHLHS (RDeH), which we have named RDeH-Net. Briefly, the framework implements a coarse-to-fine two-stage detection approach, with a structure classification network for 4D human foetal cardiac ultrasound images from various spatial and temporal domains, and a fine detection module with weakly-supervised localisation for high-precision nidus localisation and physician assistance. The experiments extensively compare the authors’ network with other state-of-the-art methods on a 4D human foetal cardiac ultrasound image dataset and show two main benefits: (1) it achieved superior average accuracy of 99.37% on three categories of foetal ultrasound images from different cases; (2) it demonstrates visually fine detection performance with weakly supervised localisation. This framework could be used to accelerate the diagnosis of eHLHS, and hence significantly lessen reliance on experienced medical physicians.