Segmentation,Feature Extraction And Counting Of Mitotic Cells In Digital Pathological Images Of Breast Cancer

Breast cancer is one of the highest cancer incidence rate and death rate in the world.According to the latest Global Cancer Statistics 2018(GLOBOCAN 2018)survey,the incidence rate and mortality rate of the three most common cancers in the world are lung cancer,breast cancer and colorectal cancer.Early detection and diagnosis of breast cancer is the key to improve the curative effect,and histological classification and grading of invasive breast cancer is an important prognostic standard,which can be used as an independent index to judge the prognosis of breast cancer.Nottingham grading systems which includes three morphological features: nuclear pleomorphism,tubule formation and mitotic count is the most commonly grading system of invasive breast cancer.Because the spread of cancer is mainly controlled by cell division,the number of mitoses is an important parameter in breast cancer grading.However,mitotic count manually is a very tedious and time-consuming work for pathologists.In addition,there are cells which will interfere with the mitotic count.these cells,such as the presence of apoptotic cells,lymphocytes and compressed nuclei,are similar to mitotic cells.After the appearance of pathological slide scanner and the popularity of deep learning method,more and more people pay more and more attention to the automatic detection of mitotic count.In the past,the methods of mitotic classification and counting of breast cancer digital pathological images were mostly based on CNN or multiple neural networks.These methods can achieve the classification and counting of mitotic cells,and they can get high recall rate and accuracy rate.However,for CNN and other neural networks,a large number of data are needed for training.As is Known to all,for histopathology,the data is very scarce.Therefore,we propose to use U-Net for mitotic cell segmentation,which is improved by full convolution network.It include data enhancement,which can solve the problem of data shortage.We choose ICPR12 as the training set of U-Net.According to the markers ofpathologists,we first convert the original images in ICPR12 to binary images,and then we use them for U-Net training.After U-Net prediction,we get the mitotic probability map,in which there are a lot of false positives.In order to reduce the number of false positives,we use the Xception network to classify the candidates as mitosis and non-mitosis,and use mitotic features to screen them.First,we obtain the central coordinates of each mitotic candidate from the probability map.Then according to the central coordinates,we extract the patches whose size are 128128?.After that,we will select some patches to train Xception.And we use the remaining patches for classfication by trained Xception network.After that,we extract the mitotic features of ICPR12 to get the mitotic area threshold.As to further reduce the false-positive rate and improve the mitotic accuracy and F-measure value,we classify the results of Xception network by the area threshold.We validate the proposed method based on ICPR14 and AMIDA13.The experimental results show that the proposed method successfully realizes the automatic segmentation,feature extraction and counting of mitosis.At last,the accuracy of the whole experiments is71.1%,which is obviously better than other experimental methods.