Research On Mass Detection Method Based On Multi-view Of Breast

Mammography is a widely used method for examining breast cancer.Clinically,radiologists often need to compare multiple views.To study multi-view based mass detection methods can improve the accuracy of detection and diagnosis,and can reduce the workload of doctors.The multi-view mass detection method is based on single-view mass detection.Firstly,Threshold segmentation and morphological methods is used to preprocess data in single-view mass detection,and nipple region is suppressed.Then the Gauss-Laplacian operator is used to extract edge informations of region of ROI.Finally,the Hessian matrix theory is used to detect bright spots with informations of edge,and then extract a collection of candidate masses.Two algorithms are used to judge whether candidate masses are false positive regions:(1)Two thresholds are found based on the gray distribution characteristics of breast.The two thresholds divides the breast ROI into three sub-regions: a fat region,a gland region and a transition region.The region where the candidate mass is located is discriminated: if the candidate mass is located near the fat region,it is determined to be a false positive region;if the candidate mass is located inside the gland region or inside the transition region,it is determined to be a false positive region.(2)Similarity matching is performed on the candidate mass using an adaptive Gaussian template,and the candidate mass with low similarity is determined as a false positive region.The Multi-view false positive detection is based on results of the single-view,mainly for views from the same perspective of both breasts.The algorithm uses three obtained breast anatomical features(nipple,chest wall line and skin line)to establish a normalized coordinate system of breast region,which enables position mapping of both views.For a pair of symmetry regions,the local binary model operator is used to extract texture features of them.If their texture features are similar,they are considered as false positive masses.There are 71 CC views and 84 MLO views in the experimental data,which are provided by the project cooperation unit.The sensitivity of CC view is 69.23%,the average number of false positives per image is 3.88,and the sensitivity of MLO view is 63.54%.The average number of false positives per image is 3.45.