Selective Search and Sequential Detection for Standard Plane Localization in Ultrasound

Abstract

We present the first automatic solution for localizing fetal abdominal standard plane (FASP) in consecutive 2D ultrasound images. FASP is located in the presence of three key anatomies detected by learning based algorithms, including stomach bubble (SB), umbilical vein (UV), and spine (SP). Traditional detection methods exhaustively scanning the entire image with sliding window algorithms tend not to perform well, since large numbers of regions appear similar to key anatomies. We propose a novel approach by applying local detectors sequentially on the preselected locations of SB, SP and UV. Specifically, we employ segmentation to generate probable locations for SB detection while exploiting a novel accumulative vessel probability algorithm to produce probable locations for SP and UV detection. The sequential scheme automatically excludes detected regions in former steps for subsequent detection, and further limits the search range according to the geometric relationship among anatomies. Experimental results on 100 fetal abdomen videos show that our method significantly outperforms traditional methods that only use local detector.