Study On Detection Methods For Microcalcification Clusters In Dense Mammograms Based On Relevance Vector Machine

Breast cancer has become the first killer of women’s health.Timely detection and early treatment of the breast cancer can effectively reduce the morbidity and mortality.At present,Mammography diagnosis is considered to be the most reliable and effective method for early diagnosis of breast cancer.Radiologists make diagnosis by observing the presence of microcalcification clusters of breast cancer in mammograms.Automatic detection of the microcalcification clusters of breast cancer using computer technology can effectively assist radiologists to improve the accuracy and efficiency of breast cancer screening.Most of the current microcalcification clusters detection methods are based on foreign non-dense mammography.The complex tissue in dense mammography makes the image contrast become low and the current detection method has to be improved to detect the microcalcification clusters.Relevance Vector Machine(RVM)is a sparse Bayesian learning model based on probabilistic prediction combined with Markov’s theory,Bayesian principle,Automatic relevance determination priori and Maximum likelihood theory.It has the advantage of being able to provide probabilistic prediction and arbitrary use of kernel functions.On the basis of Relevance Vector Machine,this thesis proposes an improved Relevance Vector Machine model based on rotation method,which improves the detection accuracy of microcalcification cluster in dense mammography.The main work and innovation of the thesis are as follows:(1)Aiming at the problem of limited number of microcalcification and insufficient learning in the training process of Relevance Vector Machine model,an improved Relevance Vector Machine model based on rotation method is proposed in this thesis.The model is used to optimize the parameters of the kernel function by cross validation method.After analyzing the sample characteristics in training samples and applying the rotation method to the relevance vector.By adding the rotation relevance vector to the original training samples,the characteristic differences and the number of training samples are increased.(2)In order to solve the problem of low precision of microcalcification clusters detection in dense mammograms,an improved Relevance Vector Machine model is applied to the detection of microcalcification clusters in dense mammograms.In this method,the feature vectors are extracted from the mammograms that have been enhanced by high-pass filtering and other pretreatments firstly.The relevant vectors are obtained by training.A new training sample set for improved relevance vector machine model is formed by adding the appropriate rotation relevance vectors to the original training sample set.Then,dense mammography is detected by the improved Relevance Vector Machine model trained by the new training sample set.And the tested microcalcification points were examined by cluster analysis.Finally,the algorithm was analyzed by the free response receiver operating characteristic curve(FROC).The experimental results show that the improved Relevance Vector Machine model can effectively increase the detection accuracy of microcalcification clusters in dense mammograms.