Based On Multi-instance Learning Ultrasound Breast Tumor Benign And Malignant Classification

Breast cancer is currently one of the the highest incidence diseases of women and it has serious implications for the health of women. Medical ultrasound imaging plays an important role in diagnosis and treatment of breast cancer, and it has some characteristics, such as effective, safe, convenient, cheap and so on. In recent years, computer-aided diagnosis (CAD) technology that could help doctors make the right decision is increasingly becoming an important tool for diagnosis. And classification in the CAD system has been one of the main difficulties.This article will focus on ultrasound classification of benign and malignant breast tumors and make some corresponding algorithm better. In this paper, the multi-instance learning algorithm is used to improve an existing CAD system and do some different improvements like the weighted decision-making based on the Citation-kNN algorithm.1. Improvement"Fully automatic and segmentation-robust classification of breast tumors based on local texture analysis of ultrasound image".This part will improve the system"Fully automatic and segmentation-robust classification of breast tumors based on local texture analysis of ultrasound image". ROI information got by ROI segmentation in this system is crude and it needs to be proceed by filter depend on priori information. So this method doesn't suit for all situtations. This article proposed a variety of structure construction methods based on the result of crude extractation in this system. And it replace the original classification method with multiple instance learning. Original classification is based on local texture features and support vector machine (SVM) classification algorithm. It improved the scope of application and classification performance.2. locally weighted Citation-kNNThis section will improve a MIL classification algorithm Citation-kNN. Before improved, this part will detailed analysis sample distribution of feature space. And proposed three kinds of distribution characteristics: relative distance, the degree of scattered and sparse distribution. And put forward corresponding weighting method for each feature, such as weighting based on relative local or global distance, weighting based on scattered degree, Weighting based on scattered degree with label modified and Integrated weighting. This article introduced a variety of weighting method to the Citation-kNN algorithm. Classification capability of the various improved algorithm will be verified by comparing the experimental results. The classification capability with integrated weighting based on global distance and modified dispersion will be better. A variety of weighted decision-making within this section presents suitable for any of the local learning method.