Recognition Of Breast Tumor Based On Ultrasonic Image

Breast cancer is the most prevalent cancer among women. The fatality ratio is keeping rising in these years. It is valuable and meaningful to research on breast cancer's diagnosis. Currently, self-examination, mammography, and sonography are the most frequently used methods for detection of breast cancers. Sonography has been widely used for breast cancer diagnosis because of its non-invasive, real time, low cost, and safe for the patient. However, it's subjective and heavily depended on operator's experience, which leads to a high positive predictive value. That means large numbers of unnecessary biopsies, which are painful and economical burdens to the patients. How to reduce breast biopsies and improve breast cancer diagnosis accuracy and objectivity has become the main challenge of computer-aided diagnosis (CAD) system. Texture feature was the most popular image feature applied while the tumor contour and gray level were less used in CAD system. This study will combine the texture, gray level and contour features in the diagnosis system to recognize breast cancer more effectively. This research focused on developing a computer-aided diagnosis system based on image analysis and image recognition. Firstly, the regions of interest were located manually. The tumor contour, gray level and texture features were extracted from the regions of intest of the ultrasonic images. Then, a feature vector was created with classification distance and artificial neural network in feature select stage. With leave-one-out method, a three-layer Back Propagation (BP) artificial neural network was applied to distinguish the malignant from the benign. The ability of computerized parameters to discriminate benign from malignant breast tumor from digitized ultrasonic images has been assessed. The images of 115 lesions, including 46 lesions, proved to malignant at histology and 69 found to be benign, were digitized and characterized by contour, gray level and texture parameters. Then, four features, namely length-width ratio, mean intensity ratio, entropy and inverse difference moment, were selected from those parameters by the classification distance and artificial neural network. Finally, BP neural network classification yielded a specificity of 46.38% while sensitivity is 100%. According to the result, 46.38% benign biopsies can be avoided without malignant misdiagnosis. With the application of the combination of tumor contour, gray level, texture features extracted from ultrasonic image, the computer-aided diagnosis system got a promising recognition result. For further research, more information such as clustered micro-calcification, flow information will be added to this system to achieve a better performance in early detection of breast cancer.