A Preliminary Study On Diagnosis Of Benign And Malignant Calcification Lesions Of Mammography And Evaluation Of BI-RADS Categories Based On Deep Convolutional Feature

Currently, breast cancer has become the most common form of malignant tumors among women. Both in developed and developing countries, the incidence of breast cancer shows an upward tendency annually. Mammography is the main imaging methods for screening and diagnosis of breast cancer, particularly sensitive for microcalcifications. Many studies have shown that mammography has high diagnostic value for breast lesions, but can not be ignored that has a high false positive ratio leading to unnecessary surgery, or a high false negative ratio resulting in delaying treatment or affecting prognosis. The paper collected a certain number of mammography calcificaiton lesions, combined with BI-RADS categories and clinical practice, to summarize the diagnostic value of mammography in the diagnosis of calcification lesions, and to explore the strategy of breast cancer screening and diagnosis.In recent years, the rapid development of computer-aided diagnosis also adds weight to imaging diagnosis of breast disease, however the level of promotion and application is far from enough due to its limitations and deficiencies. While deep learning can expand the reasearch direction. It has achieved great success in many fields, including speech and image recognition, but less involved in the field of medical image processing and identification. In this study, we try to apply the DCNNs to the image processing of mammography to extract deep convolutional feature combined with SVM classifier training which can evaluate the effect of classifier on identificaiton of benign and malignant breast calcification lesions and BI-RADS categories. The text can be divided into the following four chapters:Chapter one. The introduction. To introduce the background, research significance of the paper and summarize the strengths and limitations in breast cancer screening and diagnosis of mammography. It also narrate the present situation of computer aided diagnosis. Besides, it introduces the concept of deep learning, and elaborate on the research and achievements in medical image processing field. Finally, we relate deep learning(CNNs) to breast calcificaiton lesions of mammography, and evaluate the effect of classifier on identificaiton of benign and malignant breast calcification lesions and BI-RADS categories.Chapter two. The overview of deep learning. The main content include the basic concept of deep learning, the common deep learning methods that provide theoretical basis for the following study.Chapter three. Comparative study of mammography diagnosis and pathologic features in breast calcification lesions. This chapter is mainly to collect mammography calcification lesions with inclusion criteria combined with BI-RADS categories and clinical practice(defined Ⅳa category as positive/positive) which to summarize the diagnostic value of mammography in breast calcification lesions and correlation with pathological findings since the PACS system was running in our hospital. On the one hand, it can analyse its diagnostic value of mammography through the accuracy rate, sensitivity, specificity and ROC curve etc that make a preliminary research on breast cancer screening and diagnosis strategy. On the other hand, it also provides the basis for preliminary evaluation of classifier on identificaiton of benign and malignant breast calcification lesions and BI-RADS categories based on deep convolutional feature.A total of 383 breast calcification cases met the inclusion criteria. The diagnostic results of BI-RADS category were as follows(category 0, from category II to V):2,8,47,185 (IVa 94 cases, IVb 45 cases, IVc 46 cases) and 141 cases, and the diagnostic accuracy rates in proper order were respectively(category 0, from category ⅡI to V):50%,100%,93.62%,43.62%(IVa defined as positive) or 56.38%(IVa defined as negative),68.89%,84.78% and 99.29%; When category Ⅳ a was defined as positive disease, the diagnostic sensitivity, specificity, positive predictive value and negative predictive value of mammography in breast calcificaiton lesions were 98.82%,40.63%,76.83% and 94.55%. The area under the ROC curve was 0.697±0.022; When Ⅳa was defined as negative disease, thess same indicator results were as follows:82.75%,82.03%,90.17% and 70.47%. The area under the ROC curve was 0.824±0.021.Chapter four. A preliminary study on diagnosis of benign and malignant calcification lesions of breast and evaluation of BI-RADS categories based on deep convolutional feature. This section is mainly apply the DCNNs which be trained with ImageNet data set to extract deep convolutional feature of mammography, and combined with SVM classifier training so that it can evaluate the effect of classifier on identificaiton of benign and malignant breast calcification lesions and BI-RADS categories through collecting the mammography images which met the inclusion criteria. The results are evaluated by 10-fold cross validation.In this paper, the extraction of deep convolutional feature and SVM classifier training were conducted in the MATLAB environment, it needs more unified, standard input and corresponding output results, so it is necessary to unify the data of breast calcification lesions. The tagging for the breast calcification not only includes calcification lesions, but also reduce the interference of non-breast region and breast tissue. To identify the benign and malignant breast calcification lesions with SVM classifier, the accuracy of CC-view and MLO-view mammography calcification lesions were approximately 68.8% and 67.4%, and the area under ROC curve are respectively 0.766,0.741. It has a certain accuracy, but generally speaking the result of accuracy is not high. Besides, the classifier training effect of BI-RADS is poor, because there is a large range of overlap regions among BI-RADS Ⅳ categories(including from IVa to IVc category). The sample size of the calcification lesions of mammography is not enough to get a good recognition model for classification of BI-RADS.