Pathological Image Processing Based On Deep Learning:Breast Cancer Diagnosis And Analysis

Cancer has become a major cause of human death and breast cancer has become the highest incidence of cancer in women.Early diagnosis and treatment is an effective way to reduce breast cancer morality.However,as the "gold standard" for breast cancer diagnosis,pathological diagnosis has some problems,such as lack of medical resource and low accuracy in artificial diagnosis process.Intelligent assistant diagnosis of breast cancer provides an effective way to solve the above problems.The digitization of pathology and the continuous development of artificial neural network provide data foundations and tools for assistant diagnosis of breast cancer.This paper focuses on the intelligent assistant diagnosis of breast cancer pathological slide images,including the diagnosis of microscopic images and cancer region segmentation of WSIs.Based on the study of the characteristics of pathological slide image,investigation of the related work and analysis of existing problems,the image processing and analysis of pathological slide image is proposed for the computeraided diagnosis of breast cancer.More specifically,the main contributions of this paper are presented as follows:(1)A deep convolution neural network(CNN)for high resolution pathological microscopic images in the case of small sample data is proposed for the fourclassification diagnosis of breast cancer pathological microscopic images.A variety of data augmentation methods suitable for pathological images are applied.The knowledge from natural image is transferred to pathological images.Resnet-152 is selected to classify the images as the neural network.Experiments show that the network could be effectively trained using above methods in small samples.This model achieved an accuracy of 86% for Breast Cancer Histology Image(BACH2018)microscopic image dataset.(2)The model is further improved by using stain normalization and rescale method based on the classification algorithm in small samples.Reinhard color transformation is used to stain normalization and rescale is used to increase the actual physical area of pictures.Experiments show that the stain standardization and rescale can further improve the accuracy of the model,and the accuracy on the test set is improved to 92%,outperforming other results in BACH2018 microscopic image dataset.(3)An ultra-high resolution whole slide images(WSIs)segmentation method transferred from high resolution microscopic image classification algorithm is proposed.A picture clipped from WSIs for patch classification has same physical area as microscopic image.Gray threshold method is applied to extract the tissue area and increase the proportion of effective samples.Weighted cross-entropy loss function is used to solve the problem of data imbalance.Experiments show that microscopic image classification algorithm can be transferred in the case of different task.Compared with other related studies,the model achieved more competitive performance,with the official evaluation score of 0.6373 on BACH2018 WSI data set.The model method proposed in this paper solves the training problem of model in small sample data sets,and also proves the effectiveness of method transferred between different task and same disease.