Research On Breast Histopathological Image Classification Based On Deep Learning

Breast cancer is the most common cancer in the world,and its morbidity and mortality rate remain high,seriously endangering women’s health.Histopathological image analysis is considered as the "golden standard" when it comes to diagnosing breast cancer,but the histopathological image itself is very complex and rich in diversity,which makes the diagnosing process time-consuming,labor-intensive and inefficient.Moreover,pathologists differ in experiences,and their subjectivity may even lead to misdiagnosis.At present,deep learning has emerged in the fields of computer vision and image processing,and it also provides a new way of thinking and approach for computer-aided diagnosis.Based on the hematoxylin-eosin-stained breast histopathological image data set,this thesis studies the classification of breast pathological images based on deep learning.The main contents are as follows:Bilinear Convolutional Neural Network(B-CNN)lacks diversity in the receptive field,and the classification accuracy is not high,aiming at these problems,an improved B-CNN is proposed to automatically classify the pathological images of breast tissue in the Brea KHis dataset.Imitating the multi-scale analysis method used by pathologists in analyzing pathological images,multi-scale receptive fields was applied in B-CNN,so Inception network was used as feature extractors to replace the original VGG network.Aiming at the problem of staining differences in the data set images,a method for color normalization of histopathological images was used to make the images only change the color appearance without damaging the structural information.Due to the insufficient number of images in the data set and the imbalance in each type of image,the color normalized images was first flipped,and then image patches were extracted from each type of image according to a certain ratio,after that,the number of training samples for each type of image is at the same number,which solved the problem of sample imbalance and insufficient.The improved network is used to classify the histopathological images in four different magnifications,and the network is finetuned using transfer learning.The experimental results show that the improved B-CNN network improves the classification accuracy.Based on deep residual network Res Net101,the residual block of Res Net101 was improved,then SE module was added to the improved Res Net101,the SE-Res Net-B network was desiened and applied to classify the high-resolution breast histopathological images into two classifications and four classifications.Since high-resolution images cannot be directly fed into the network for training,sliding window of size 512×512 was applied to sample high-resolution images after the color normalization process to ensure that the diagnostic information of the image can be included,and the sliding overlap rate of 50% increases the continuity of features between the sampling patches.In the classification strategy,firstly the sampled image patches was classified,and then the majority voting algorithm was used to obtain the classification results of the original high-resolution images.The experimental results show that the improved network model SE-Res Net-B significantly improved the classification performance,and achieves a binary classification accuracy of 96.25% and a four-category classification accuracy of 88.75% on the ICIAR 2018 dataset.