Complementary Network for Accurate Amniotic Fluid Segmentation From Ultrasound Images

Abstract

This study presents an automatic method for estimating antenatal amniotic fluid (AF) volume from two-dimensional ultrasound (US) images, which is an important indicator of fetal well-being. This automatic estimation of AF volume (AFV) requires automated segmentation of the AF pocket, which is a challenging task due to its amorphous features and US artifacts, such as reverberation, shadowing, particle noise, and signal dropout. Recently, AF-net, a deep-learning method, has been shown to successfully perform AF pocket segmentation. However, we observed that AF-net is prone to misjudging AF pockets containing severe reverberation artifacts. The proposed method addresses this problem by developing a dual path network, which consists of AF-net as the primary path and an auxiliary network as the secondary path. The auxiliary network is designed to focus on the local area that is likely to be contaminated with the reverberation artifacts. It infers this local region and generates a feature map of the artifacts, incorporating it as prior information into a deep neural network, denoted as RVB-net, for segmenting the reverberation-artifact-contaminated AF region. Finally, the segmentation output from the auxiliary network complements the AF-net. Experimental results show that the proposed dual path network effectively reduces misjudgment of the AF pocket caused by severe reverberation artifacts. The proposed dual path network achieved an average Dice similarity coefficient (DSC) of 0.8599 ± 0.1074 (mean ± standard deviation) for AF pocket segmentation on the entire evaluation set.