Breast Calcification Classification By Deep-learning And Radiomics Descriptors

In current clinical practice,mammography is the most convenient non-invasive detective way and plays a key role in the early stage of breast cancer diagnosis.Conventional mammography based diagnosis mainly lies on radiologists,which is always time-consuming and easily affected by multiple subjective factors.Thus,the application of computer-aided detection and diagnosis system in mammography has always been the research focus due to more reliable and precise output with pretty less time invested.Sophisticated procedures based on traditional machine learning and image processing techniques have been developed for the diagnosis of mammographic calcification.However,previous researches all suffered from high false positive rate.Characterizing the calcification in an automatic and stable way remains a challenge.Incited by the successful applications of reliable and repeatable radiomics approaches and deep learning machines,this paper made improvements as follow:Firstly,this paper makes a brief introduction of traditional diagnostic process in mammographic calcification,and the development of the convolutional neural network.Secondly,we utilize the traditional image preprocessing method to propose a novel semi-auto segmentation method.Thirdly,this paper focuses on quantifying the calcification and studying the effect of traditional radiomics approaches combined with different feature selection strategies and classification models.At last,due to the outstanding performance of convolutional neural network in natural image recognition,this paper introduces a feature selection method and a visualization method for deep descriptors.We trained several deep learning models to ensure the effectiveness of the proposed strategy.Experimental evaluations in Sun Yat-sen University Cancer Center demonstrate the effectiveness of our approach.In conclusion,the main contributions are as follow:(1)This paper analyzes traditional diagnosis process,and then utilizes fundamental segmentation methods to propose a semi-automatic segmentation method for radiologists to segment calcification.(2)Depending on the radiologists' considerations about the distribution and the density of calcification cluster,this paper invents new features to describe them and analyze the diagnostic performance for the radiomics descriptors by the combination of varies feature selectionmethods and classification methods.(3)To avoid the over-fitting,we apply the proposed feature selection method in deep learning features guided by radiomics descriptors.At the same time,we try to visualize and explain the selected neurons to analyze the clinical value of deep learning methods.