Automatic Segmentation of Abdominal Subcutaneous Adipose Layer in Ultrasound Image Using CNN

Accurate measurement of the thickness of subcutaneous adipose tissue (SAT) can effectively estimate body composition. Ultrasound is an accurate technique for measuring the thickness of SAT layer. However, at different body sites, the SAT layer has different content of inlaid fibrous structure which causes segmentation of the SAT layer to be very difficult. This paper presents a fully automatic approach to detect and extract the border of the SAT layer using Convolutional Neural Network (CNN) techniques. Our approach utilizes CNN to learn the complex regression function that maps the borders into their positions in images. The SAT layer is segmented according to the predicted upper and lower borders and its thickness is calculated automatically. The average predicted error on simulated ultrasound images achieves 0.34mm or 1.88% of the SAT thickness. On tested human abdominal ultrasound images, we obtain the average prediction error of 0.85mm or 8.71% of the SAT thickness. This study is suitable for segmenting subcutaneous fat tissue, and it may be used as a general framework for applications of regression CNN to identify other specific borders on ultrasound images. Introduction Body composition has a large impact on health and physical performance associated with excessively high or low amounts of body fat [1]. Overweight and obesity are becoming major health problems of the 21st century in more and more countries [2]. The thickness of subcutaneous adipose tissue (SAT) is accepted as a body fat indicator because about 40 to 60% of total body fat is in the subcutaneous regions [3]. Moreover, the detection of the thickness of the SAT at rectus abdominis is more important, because it is closely located to healthy critical abdomen organs. Clodagh et al. [4] proposed an ultrasound scanning protocol for subcutaneous fat and measured subcutaneous fat thickness using B-mode images and showed that ultrasound is a reliable, reproducible, accurate and safe method for measuring subcutaneous fat as well as muscle thickness. Müller et al. [5,6,7] proposed a semi-automatic method with a region-growing algorithm for detecting SAT layer on ultrasound images and this method would select possibly the wrong region of interest (ROI) owing to the embedded fascia of the subcutaneous fat. Ng el at. [8] presented an approach to measure subcutaneous fat thickness automatically by ultrasound radio frequency (RF) signals instead of ultrasound images. They segmented fat boundary at the suprailiac, triceps, and thigh sites using the spectrum dispersion of the power spectrum of RF signals. However, complex structure with much fibrous connective tissue of rectus abdominis was not investigated in their paper. Recently, convolutional neural networks (CNN) have been widely used in the area of the classification, segmentation, object detection and registration on medical images [9]. In this paper, we propose a novel fully automatic approach using linear and quadratic regression functions of CNN models to extract the upper and lower borders of the SAT layer. The training set of the CNN is a set of ultrasound images where the locations of the SAT borders are labeled manually. SAT borders defined as the upper and lower boundaries can be approximated as a straight line and a