Multi-level trainable segmentation for measuring gestational and yolk sacs from ultrasound images

Abstract

As a non-hazardous and non-invasive approach to medical diagnostic imaging, ultrasound serves as an ideal candidate for tracking and monitoring pregnancy development. One critical assessment during the first trimester of the pregnancy is the size measurements of the Gestation Sac (GS) and the Yolk Sac (YS) from ultrasound images. Such measurements tend to give a strong indication on the viability of the pregnancy. This paper proposes a novel multi-level trainable segmentation method to achieve three objectives in the following order: (1) segmenting and measuring the GS, (2) automatically identifying the stage of pregnancy, and (3) segmenting and measuring the YS. The first level segmentation employs a trainable segmentation technique based on the histogram of oriented gradients to segment the GS and estimate its size. This is then followed by an automatic identification of the pregnancy stage based on histogram analysis of the content of the segmented GS. The second level segmentation is used after that to detect the YS and extract its relevant size measurements. A trained neural network classifier is employed to perform the segmentation at both levels. The effectiveness of the proposed solution has been evaluated by comparing the automatic size measurements of the GS and YS against the ones obtained gynaecologist. Experimental results on 199 ultrasound images demonstrate the effectiveness of the proposal in producing accurate measurements as well as identifying the correct stage of pregnancy.