Research On Diagnosis Methods Of Breast Masses Based On Reference Images

Computer-aided diagnosis(CAD) of breast masses based on reference image, not only provides the decision value of the suspicious mass of breast image for diagnosis, which indicates its probability of being a mass, but also provides a series of diagnosed reference images which are similar to the image for diagnosis and compared by doctors. Compared with the conventional CAD of breast masses which only provides decision value, it has unique advantages.The research focuses on CAD of breast masses based on reference images. The main contents include segmentation of suspicious mass, feature selection and calculation of feature weight in measurement of similarity, calculation of decision value. For segmentation of suspicious mass, an improved dynamic programming based segmentation algorithm is proposed. At first, it confirms the coverage region of valid edge points by using the method based on template matching, which can eliminates the interferences coming from some not-edge tissues with strong edge characteristic. Then the weight of each component of edge cost is calculated by chaos particle swarm optimization algorithm. It's easier to get appropriate weight than manual setting. On feature selection and calculation of feature weight in measurement of similarity, the feature weight is calculated by the particle swarm optimization algorithm which includes a kind of off-trap strategy. The feature will not be selected to measure similarity when its weight is zero. The method can integrate feature selection and calculation of feature weight into one process. For calculation of decision value, a hybrid classifier to calculate the decision value of suspicious mass is proposed. Combining KNN classifier and logistic regression classifier, it makes full use of the advantage of each classifier under certain conditions.Using 210 breast images with gold standard of mass edge to test segmentation, the result shows the improved algorithm improves the segmentation performance. Using 1220 breast images to test calculation of decision value, the Az value of ROC curve of hybrid classifier is 0.8842, which is better than that of KNN classifier and the logistic regression classifier. It shows that the hybrid classifier can improve classification of decision value.