Breast Tumor Ultrasound Image Segmentation Algorithm Research And Application Of Dynamic Contour

As the key step in the computer aided diagnosis system for breast tumors,the automatic segmentation of the tumor region from ultrasound images has the great clinic significance.However,the segmentation is difficult to implement due to the low quality of ultrasound images,such as the heavy speckle noise,the low image contrast and the intensity inhomogeneity.The active contour model(ACM) is a novel method for the image segmentation and has shown its potential in the medical image processing with its unparalleled merits.This dissertation mainly focuses on the research of the level-set based ACM for the automatic segmentation of ultrasound breast tumor images.Existing ACM algorithms encounter certain drawbacks when applying to ultrasound images,such as their sensitivity to the noise and initial contours,and mis-segmentation under the intensity inhomogeneity.In this dissertation,three effective segmentation methods are proposed for ultrasound images.For images with the intensity inhomogeneity,the first method emphasizes on the local intensity information and can extract accurate tumor boundaries.However,this method is computation-expensive,thus the second one makes improvements and reduces the computation burden to a certain extent.Inspired by the coarse-to-fine analysis strategy in the multi-resolution theory,the third model is able to exploit features of the image at different resolutions and accelerate the processing speed.Moreover,the combined region- and edge-based information in this method can offer a considerable robustness and accuracy in the segmentation task.Based on extracted boundaries,a computer aided diagnosis(CAD) system is established for the automatical classification of breast tumors.This system utilizes 5 morphologic features and 3 texture features as inputs of the classifier.To train the classifier(an artifical neural network as in this dissertation),a novel selective ensemble learning algorithm is proposed based on pseudoinverse,which is easy to implement and efficient in the computation complexity,and moreover,can achieve a high generalization ability.Experiments on 132 pathologically proven cases including 68 benign tumors and 64 malignant ones show that the proposed system yields a classification accuracy of 88.91%。...