The Research Of Automatic Segmentation Based On Object Recognition For Breast Lesions In Ultrasound Images

Breast cancer is one of the common tumors in women. Due to its substantialmorbidity and high risk, more and more medical researchers have been watched closelyof the breast cancer. How to make an accurate diagnosis of the breast tumor patients isthe key to the prevention and timely treatment of breast cancer. Ultrasound has provento be an important technological for the diagnosis of the breast cancer, which has theadvantages of no radiation, easy availability and low cost. However, with thedifference of the experience and theoretical, it is hared to obtain an unified diagnosistowards the same ultrasound image between clinicians. With the develpment ofcomputer science and medical imaging technology, computer-aided diagnosis hasplayed an unreplacable roles in mediacal image analysis, lesion area detection, tumorrecognition and providing medical advice. But the specular nature of sonogramsproduces some speckle noises and low constrast in the ultrasound images, whichpresent difficulties in computer-aided diagnosis.This paper introduces a novel automatic segmentation method of breast lesions inultrasound image, which incorporates a robust graph based segmentation and an objectrecognition method. Traditional segmentation methods of ultrasound image mainlyfocus on the basic function of segmentation, and the follow-up work is heiglyoperator-dependent. In order to reduce the influence of different experience and relieveworking stress in clinicians, this paper proposes an novel automatic segmentationscheme based on an object recognition method for the segmentation of breast leisionsin ultrasound image. In this scheme, a total-variation method is first applied to reducethe speckle noise in the ultrasound image. A robust graph-based segmentation methodis then used to segment the image into a number of sub-regions. An object recognitionmethod which incorporates image feature extraction, feature selection andclassification is proposed to detect the breast tumors from all the sub-regionsautomatically. Finally, an active contour model is used to refine the contours of theregions that are recognized as tumors.In our experimence,46breast ultraound images with diagnosed tumors are used tovalidate the proposed scheme, in which23are diagnosised to be benign and theremaining are malignant. The experiment results show that the accuracy, specificityand sensitivity of recognition are98.3%,98.5%and97.4respectively, and the TP, FP,FN and ARE of segmentation are85.01%,1.78%,14.99%and9.08%respectively, validating that the proposed scheme is good and acceptable in automatic segmentatin ofthe breast lesions in ultrasound image.