Segmentation of Breast Ultrasound Lesion Boundary using Texturebased Multi-resolution Method

摘要

Computer-aided characterization of a breast ultrasound lesion involves two steps: first, extracting features from thelesion whose boundary is pre-defined on the images, and then converting the features into mathematical models. Mostmethods assume that the boundaries of the lesions are pre-selected or outlined by sonographers or operators, becauseautomated delineation of lesion boundary is not trivial and is a challenging task. The purpose of this study was todevelop and evaluate an automated lesion boundary segmentation method that is based on texture-based, multiresolutionimage analysis. One hundred ninety-seven breast ultrasound images containing solid breast lesions from 172women (age 24-89 years, mean 38 years) were studied. Fifty-five of the 197 images were from 40 women withmalignant lesions, and the remaining 142 were from 132 patients with benign lesions. Each breast lesion was identifiedby an operator who placed a rectangular region of interest (ROI) to widely encompass the lesion. The resolution of theimage was compressed, at variable ratios depending on the ROI size, to reduce noise. Texture momentum wascomputed. A binary image was generated from the texture and pixel intensity parameters. Initial seed boundary wassegmented from the binary image and then expanded to the original resolution using the boundary pixel intensitygradient information. The boundary of each breast lesion was delineated by a breast-imaging radiologist who wasblinded to the computer-detected lesion boundary. The ‘area match ratio’ between the manually drawn boundaries andthe automatically detected boundaries was computed. This ratio is equal to or less than unity (unity indicates that theareas match exactly). Overall, good agreement was seen between the multi-resolution segmentation method and theradiologist’s manual delineation. The mean area match ratio was 0.87 ±0.02. We have developed a multi-resolution,texture-based method to segment the boundary of breast lesions. This method will facilitate full automation for thecharacterization of breast ultrasound lesions.