The Research On Classification Of Malignant And Benign Clustered Microcalcifications In Mammography

Breast cancer is one of the common disease in female.The micro-calcifications cluster(MCC)in the mammography which is a early symptoms of breast cancer can be used for early diagnosis.MC detection in the early stage is one the key technologies in breast cancer diagnosis.In order to help the doctor to find the breast cancer in the early stage,a classification of malignant and benign clustered MCC based on geometry and texture feature by using generalized regression neural network is presented in this thesis.We investigate the geometry texture of MC and MCC.By union those useful feature mentioned in the above,we proposed a innovative method for the classification of MCC.The major work of this thesis is focused on the assessment of the MCC.And achievements are shown as below:(1)MCC region extraction:We proposed a new distance based method for MCC extraction in this thesis.By employing this method,a convex hull region is extracted from the ROI.This method can subtract the useless region while keeping all useful MC information which is a very import step for improving the ratio of correct diagnosis in the fellow step.(2)Taking many factors into consideration in the practical application as the detection of MC may have some error ratio.the ROI for the MC diagnosis in image processing step are not of uniform size.In order to deal with those problems and improve accuracy.Beside the geometry texture,it is supplemented by image texture.We add GLCM and wavelet coefficient to improve the diagnosis.Several feature selection methods are employed to select the best feature subset.(3)In order to build an effective model for classification of MCC with a limit number of training sample data.We adopt the generalized regression neural network to build the training model.Meanwhile in order to handle the case that there is micro-calcifications in the ROI but not a valid cluster is presented in the image,a new method by using Fuzzy C means cluster center for default feature evaluation of no MCC ROI is represented in this thesis.So we can remain the MC geometry texture from ROI for training while retaining the number of sample data and improving the diagnosis accuracy.The experimental result shows that the method based on the image texture of MCC has better practicability.This method can be used for MC ROI image diagnosis in a real application when the MC is not accurately segmented.A new method from this thesis is put forward for breast cancer CAD detection.