|
Ultrasound imaging technology is currently the most commonly used means of early detection of breast tumors one. Ultrasound can be real-time imaging, the price is relatively inexpensive, easy to use and noninvasive, is of great significance in the clinical diagnostic imaging. High-performance computer-aided diagnosis (Computer Aided Diaganosis, CAD) system, can further improve the accuracy of detection of breast tumors, and doctors to provide effective clinical diagnostic suggestions. However, taking into account the ultrasound images of breast tumors serious speckle noise, artifacts, low image contrast, the shape of the tumor may be complex and changeable characteristics, automatic detection of the tumor region and benign and malignant lesions classification is still more difficult. The goal of this paper is to develop a new, efficient method of computer-aided diagnosis, the premise without human intervention, to achieve automatic detection of the region of interest in ultrasound images of the breast tumor, tumor edge extraction as well as benign and malignant classification. The first portion, based on the texture feature of the pixel point classification, can automatically identify the region of interest (Region of Interest, the ROI). First, to remove speckle noise breast tumor after the ultrasound image is divided into non-overlapping sub-blocks, each sub-block, respectively, to calculate the local texture features, characteristics and location characteristics of the local gray level co-occurrence matrix, and synthesis of the above features to describe each sub-block . Secondly, the extracted features, feature selection, the selected feature set as a self-organizing map neural network classifier input vector, you can distinguish between the original ultrasound image of the candidate ROI region and background region. Finally, based on the characteristics of ultrasonic images mammary tumor size, location, distribution, etc., removal has a similar texture characteristics of the impact of pseudo-ROI area, the final ROI automatic detection result can be obtained. In this thesis can automatically identify the ROI, segmentation for follow-up to the edge of avoid doctors handmade framing workload. The second part, a band weighting neighborhood grayscale information Normalized Cut (Ncut) method, automatic extraction of breast ultrasound images of the tumor edge. First, Ncut breast ultrasound image sub-block, and the block of gray scale and space distribution characteristics to identify the contours of the tumor to be detected. Secondly, the minority split inaccurate results (such as: gray-scale leak), the choice of combination of local energy term dynamic contour model of the Ncut extract the initial edge correction, to make it more realistic target contour. Compared to a conventional edge extraction algorithm, this thesis method calculation amount is small, in the case without any manual intervention (such as: artificial initial contour) accurately and efficiently implement automatic segmentation of the tumor, and thus is expected to further improve the computer-aided diagnosis The degree of automation. The third part, through the analysis of benign and malignant tumors showed the difference in the ultrasound image, combined with the clinical experience of doctors extracted from a group of acquisition system does not depend, robustness, characterized recognition ability to describe various tumor here were collected Three texture features five morphological features. On this basis, the paper introduces efficient Affinity Propagation (AP) clustering method, as the classification of benign and malignant tumors discrimination, proven method in the premise of no training process to achieve the existing database of benign accurately distinguish malignant. In addition, the paper analyzes and compares the AP clustering method and neural network tools (such as: back-propagation artificial neural network, Fisher linear discriminant generalization capability of support vector machine) to verify the performance of various classification methods . Based on the above three parts to build the system of computer-aided diagnosis of breast cancer, 132 cases collected clinical ultrasound images of the breast tumors (including benign 67 cases, of vicious 65 cases) test, and paper method performance with the traditional level set edge segmentation algorithm, individual artificial neural network classifiers compare. The experimental results show that the system proposed in this paper has a high accuracy of ultrasound images in the detection of breast tumors and benign and malignant discrimination, computation, no training process, a high degree of automation. Thus, the system is expected to provide valuable suggestions for the physician's clinical diagnosis.
|