Surface Extraction Using SVM-Based Texture Classification for 3D Fetal Ultrasound Imaging

Abstract

This paper presents a new method for extracting the frontal surface of a fetus automatically from a three-dimensional (3D) fetal ultrasound volume using support vector machine (SVM) based texture classification. Since a fetus often floats on amniotic fluids in its mother's uterus, the major part of the frontal surface may be extracted removing dark regions corresponding to the amniotic fluid regions. In this method, the removal of dark regions in a VOI of the volume is performed by a Laplacian-of-Gaussian (LoG) followed by zero-crossing detection, which is called coarse segmentation. In the regions segmented coarsely, some are fetus regions, some non-fetus regions such as the uterus, abdomen, and floating matters, and other mixed ones of the two. In order to extract more pure fetus regions, fine segmentation is executed to split the regions into more homogeneous sub-regions. The textureness of each sub-region is then measured by multi-window BDIP and multi-window BVLC moments and classified into fetus and non-fetus ones by a SVM which is known as efficient classification tool. The frontal contours extracted from merging adjacent fetus sub-regions is combined in all the frames of the VOI to generate a fetal surface, which defines a mask volume for 3D visualization of the fetus. Experimental results show that the proposed method is useful for automatic visualization of a fetus without intervention of a user in 3D ultrasound imaging.