| Breast cancer has become one of the most dangerous malignant tumors to women’s health.At present,the diagnosis of breast cancer is mainly based on artificial microscopic analysis of breast biopsy,which is not only time-consuming and laborious,but also subjective and differential.Therefore,if the diseased region can be automatically and accurately diagnosed by computer,it will not only save time and effort,but also provide accurate and objective evaluation basis for pathologists,which is of great significance for the intelligent diagnosis of malignant tumors.This can be achieved by semantic segmentation the of the slice image of breast tissue according to the canceration or not.The image segmentation of breast tissue biopsy whole-slide image studied in this paper is based on pixel-level segmentation with Gigabit resolution.Because of the huge amount of data,complex background and different target scales,the existing image segmentation algorithms can not be well applied to the data in this paper.Aiming at some difficulties in semantics segmentation of the whole-slide images of breast cancer,combining with the latest in-depth learning technology,an image segmentation algorithm combining multi-scale atrous convolutional neural network and deformable model is proposed for multi-classification of breast cancer(normal,benign,malignant(in situ or invasive)):(1)This paper adopts multi-scale atrous convolution structure obtains multi-scale features for feature fusion and uses deconvolution to reduce feature loss in the upsampling to adapt to more refined multi-classification tasks.This improved multi-scale atrous convolution neural network is trained and validated on the data set.Compared with the classical algorithm,the accuracy of this method is higher.(2)The convolution neural network adopts the method of random extraction of image blocks in training,which makes the network unable to obtain a wider range of organizational characteristics and global information,and may lead to the lack of contour details inbreast tissue segmentation.In order to optimize the results of network segmentation,a deformable model integrating global shape information and prior knowledge is proposed in this paper.Experiments show that the above improvements have good segmentation performance in multi-class semantics segmentation of breast cancer full-scan slice images,and lay a good foundation for assistant diagnosis and related research. |
