Ultrasonic Breast Cancer Screening System Design And Application

Breast cancer is one of the leading causes of deaths from cancers for the female population worldwide. Breast Cancer Sifting Program (BCSP) can decrease the mortality rates by40%every year. So far, due to the fact of being noninvasive, portable, real-time and cost effective, ultrasonography has been recommended as the preferred imaging sifting method for early detection, instead of other approaches. However, screening each subject results in hundreds of ultrasound images. It is inefficient and fatigue for the experienced sonologists to deal with a mass of data because of the promotion of the BCSP, which might cause failure and error in the detection of lesions. In this dissertation, we make out a whole breast screening mechanism and also establish a novel and effective computer-aided sifting (CAS) system, which can help to sift abnormal breasts automatically.In the first part, based on the characteristic of the breast ultrasound dynamic images and the principle of persistence of vision, an inter-frame sub-sampling method is used to reduce the speckle noises and processing time. Breast glands region is extracted by combining Sobel edge detector with location information, because breast tumors always appear in the glands regions.In the second part, two effective methods are proposed to detect the suspicious abnormal regions (SAR). To solve the problems of the speckle noise, pseudo image. low contrast and image inhomogeneities, a method based on the improved simplified pulse coupled neural network (SPCNN) and fuzzy mutual information model is firstly proposed to detect the SAR of breast tumor ultrasound images. The ultrasound image is mapped to the fuzzy sets to enhance the contrast, then the SPCNN model is used to pulse the ultrasound image, and the fuzzy mutual information is used as the optimization criterion to obtain the relative classification results. This method improves the accuracy of SAR detection to a certain extent, however, the processing time is relatively slow. Thus a modified multi-scale normalized cut (Ncut) method is proposed with regional texture information, which makes improvements in accuracy and reduces the computation burden. Here ultrasound image is firstly partitioned into clusters with the proposed method, then the intensity and shape criterions are used to exclude the useless regions.In the third part, three intra-frame and three inter-frame system-independent. robust and effective features are defined for each suspicious region, a support vector machine (SVM) classifier is used to exclude the tumor-like non-tumors. This can reduce the false positive rates.In our preliminary application, there are480videos from40patients examined by this system. All possible tumors can be found out, while the false-positive rate can be controlled below40%. The results show that this CAS system can recognize and locate the abnormal areas from breast ultrasound images. It can improve diagnosis efficiency and objectivity in the premise of guaranteeing diagnosis accuracy. Therefore, the system may play an efficient sifting role for breast tumors from ultrasound images.