Deep Learning Based Breast Cancer Histopathological Image Intelligent Recognition

Breast cancer is a serious worldwide health problem.Breast cancer deaths in 2012 are520,000 and it is estimated that 6.2 million new cases will appear in 2017,according to reports of the world health organization?WHO?^[1].The golden standard of the final diagnosis of breast cancer is histopathological analysis,which is usually carried out by pathologists.This inspection is very time-consuming and can be influenced by various subjective factors.With the increase in new cases of breast cancer,a computer-aided detection/diagnosis system is urgently needed to assist pathologists for classification and grading.Identification of benign and malignant breast tumor and differentiation degree of mutation cells in breast cancer from pathological images with HE staining is two important steps of pathological image aided diagnosis system for breast tumor.Targeting core needs of these two kinds of breast tumor pathological imaging computer-aided system,this paper proposed multi-task benign and malignant classification method of breast cancer based on deep learning and light-level cascade convolutional neural networks segmentation frameworks from end to end.The study about benign and malignant classification method for breast cancer pathological images with HE staining mainly includes the following two aspects:first,targeting on the main reason that limits the advancement in this field is the shortage of large and open breast cancer pathological image dataset,this paper introduced a breast cancer pathological image benign and malignant classification dataset named BreastSZU_v1,which includes 15 cases,4985 breast cancer HE staining pathological images.Currently,this dataset is open to application due to research.Secondly,in order to solve the large difference in between-group results of cross validation which is resulted from staining difference in pathological images and image differences caused by different magnifications,this paper designed a multi-task classification network structure is to solve it.This network will classify color,magnification,benign and malignant of images,which can improve the network's accuracy for classifying benign and malignant breast cancer.Experiments proved that multi-task network structure has significantly improved the results of cross verification which has poor result in one-class classification network and it has also significantly reduced the between-group result difference of cross verification.About the research on differentiation degree of cancer cells in pathological imaging of breast cancer,this paper takes it as a segmentation problem of cancer cell in breast cancer HE staining pathological imaging.U-Net and FCN are the most popular deep convolutional neural networks.But in our actual experiments,U-Net and FCN have problems like high computer complexity?too much trainable parameters?,high occupation of computing resources,etc.The deficiency of this model design leads to high hardware cost,low computing speed and other problems in actual application.In order to make up the above disadvantages,this paper proposes a light Convolutional Neural Network segmentation framework.This framework consists of a segmentation network and a shape refine network.Shape refine network of cascading behind segmentation network will modify shapes according to output predicating figure of segmentation network,to make up the problem of low segmentation accuracy caused by light design of segmentation network.Besides,targeting the problem of low training efficiency caused by imbalance between object segmentation and background region in original image,this paper proposed a new objective function of weighted euclidean distance,which can force the network refresh those pixel classification tasks that cannot be classified correctly.Experiments proved that by using models trained by cascading segmentation framework and objective function,the trainable parameters of this model will only be 1/6 of U-Net if the segmentation performance of U-Net is similar,while the occupied computer resources by training and testing will only be half of U-Net.