Research Of Deep Learning Technology In Breast Pathological Image Analysis

Due to different living environments and habits,cancer,as one kind of malignant tumor,has seriously damaged people’s physical health and quality of life.Breast cancer,the most common cancer in women,is one of the leading causes of death worldwide.Pathological diagnosis,which extends visible organs to microscopic cells,has been widely considered as the ’gold standard’ of clinical diagnosis.However,the pathological image data is huge and professional.Manual tagging is time-consuming and highly subjective.Therefore,it is an important task to develop computer-aided pathology diagnosis algorithm.In recent years,with the great success of convolutional neural network in natural image processing,a large number of researchers have been inspired to use deep learning technology to develop computer-aided pathological diagnosis algorithms.However,the lack of publicly available annotated data has been identified as one of the major challenges in the application of deep learning techniques to the biomedical field.The accuracy of computer-aided diagnosis algorithm which can be accepted by clinical diagnosis needs to be further developed.In this paper,the most challenging application in current pathological image research,’detection of mitotic cells in breast cancer pathological image’,is analyzed.The characteristics and problems of current computer-aided diagnosis of pathological image are analyzed,and various deep learning and machine learning technologies are tried to realize efficient detection process of pathological image cells.The main work of this paper and the main algorithm proposed are as follows:(1)the reasons for the high false negative and false positive in previous mitotic cell detection algorithms were analyzed,and a two-stage cell detection algorithm combining multi-scale classification network and similarity distance prediction network was proposed.(2)aiming at the limitations of the mimic cells in the two-stage cell detection algorithm and the disadvantages of the traditional generative adversarial network.The triplet loss and diversity loss are introduced into the original loss function.The ability of the original two-stage cell detection model is enhanced by using the generated mimic data.(3)the algorithm in this paper is verified in the public data set.Compared with the most advanced algorithm,the algorithm proposed in this paper has the highest recall rate score and similar F-score.