Breast Cancer Classification Of Histopathological Images Based On Convolutional Network

Early detection,diagnosis and treatment of breast cancer can provide patients with a higher chance of survival.Early diagnosis of breast cancer includes X-ray imaging,mammography,MRI and other methods.Through histopathological images of breast cancer,the patient’s diseased area is further diagnosed as a cancer level,which serves as a reliable basis for subsequent treatment.Provide accurate and effective treatment plan.Pathology image classification methods have gradually changed from manual feature extraction combined with machine learning classification algorithms to deep learning classification algorithms,which has realized the improvement of pathology image classification automation and improved recognition efficiency.Compared with traditional machine learning algorithms,the convolutional neural network in deep learning has good performance in the fields of image classification,segmentation and target detection.Therefore,this article focuses on the classification of histopathological images of breast cancer,mainly using the convolutional neural network model to achieve the purpose of pathological image classification.The summary of the work is as follows:1.This article selects ICIAR2018-DATASET and Break His-DATASET small sample data sets.When using convolutional neural network model training data,small data samples will make the network unable to learn more features,and ultimately lead to poor model performance and poor generalization ability;On the other hand,due to the large size of the original image in the data set,directly inputting it into the network is likely to cause the network to collapse.Therefore,this paper designs a strategy for extracting patches,only retaining the patches with valuable feature information,and then performing contrast stretching processing on the patches to achieve the purpose of data enhancement.2.Nowadays,with the increasing arduousness of sample data processing tasks,convolutional neural networks are also showing the trend of deep-level networks,such as Res Net,Res Ne Xt network,etc.The deep network can extract deeper features in the image and improve the performance of the model.Therefore,this paper designs a deep Res Ne Xt network to classify pathological images.The final experimental results are: the accuracy rate at the patch level is 71.92%,and the accuracy rate at the image level is 82.5%.3.The breast cancer pathology image data set is composed of four categories,namely normal,benign,carcinoma in situ and invasive carcinoma.Each category is a change from the previous category,so between two adjacent categories,Their nucleus characteristics and surrounding tissue morphology have a certain similarity,which provides a challenge for the identification of similar species.For this reason,this paper designs an octave convolutional layer based on the idea of octave convolution,increases the feature extraction of highfrequency information,reduces the redundancy of low-frequency feature information,and improves the effect of extracting similar details.The improved network model For ROnet.The final experimental results are: the accuracy rate at the patch level is 81.73%,and the accuracy rate at the image level is 90%.4.Due to the improvement of the octave convolutional layer introduced to the basic network Res Ne Xt,the depth and width of the network increase,which greatly increases the training parameters of the network model and increases the computational overhead.For this reason,this paper introduces heterogeneous convolution,and designs the heterogeneous convolution layer in this paper according to the idea of heterogeneous convolution.The improved network model is ROHnet,which reduces the network model parameters under the premise of ensuring the recognition rate.The final experimental results are: the accuracy rate at the patch level is 83.04%,and the accuracy rate at the image level is 91.25%.In order to prove the performance of the final network model ROHnet,experiments were conducted on the data set Break His-DATASET,and the final result is: the accuracy rate at the image level is 95%.Finally,it shows that the network has certain robustness and generalization ability.