Complementary Network for Accurate Amniotic Fluid Segmentation From Ultrasound Images

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