| Summary: | This paper presents a novel, fully-automatic framework for
fetal echocardiography analysis of full-length routine firsttrimester fetal ultrasound scan video. In this study, a new
deep learning architecture, which considers spatio-temporal
information and spatial attention, is designed to temporally
partition ultrasound video into semantically meaningful segments. The resulting automated semantic annotation is used
to analyse cardiac examination workflow. The proposed 2D+t
convolution neural network architecture achieves an A1 accuracy of 96.37%, F1 of 95.61%, and precision of 96.18% with
21.49% fewer parameters than the smallest ResNet-based architecture. Automated deep-learning based semantic annotation of unlabelled video scans (n=250) shows a high correlation with expert cardiac annotations (ρ = 0.96, p = 0.0004),
thereby demonstrating the applicability of the proposed annotation model for echocardiography workflow analysis.
|
|---|