Research On Pathological Image Classification Of Breast Cancer Based On Deep Learning

The pathological image of breast cancer is the gold standard for the final diagnosis of breast cancer by doctors.The early diagnosis of breast cancer can effectively improve the cure rate of patients.Therefore,the classification of breast cancer pathological images has great research significance and value.Cancer cells are highly coherent,leading to subtle differences in different categories of breast cancer pathology images.Even the resolution,contrast and appearance of the same category of images are larger than the different categories,making the classification of breast cancer pathology images a huge challenge.The use of deep learning to classify breast cancer pathology images is an important research direction in the field of biomedicine.Due to the characteristics of breast cancer pathological images and the insufficient collection of publicly labeled breast cancer pathological medical image samples.As well as the difficulty in training optimization after deepening the depth of deep learning model training,it becomes difficult to classify pathological images by deep learning.In response to the above problems,this paper has done the following research:(1)A breast cancer binary classification model based on ResNet50 network is proposed.In order to avoid the network degradation problem caused by the excessive depth of the Convolutional Neural Network(CNN),the residual network ResNet50 containing the residual block is selected as the basic network framework,and the transfer learning technique is used to solve the over-fitting caused by the sample shortage.The problem is an automatic classification model for breast cancer pathology images.The experimental results show that the model achieves high precision and high sensitivity recognition performance of the data set.(2)Breast cancer binary classification model based on AutoAugment(AA)and ResNet50 network is proposed.The model is divided into two sub-modules,one of which is a search module of the automatic enhancement strategy.The sub-module uses the reinforcement learning-based search algorithm for the uncertainty of the traditional manual enhanced extended data to the model.The algorithm is controlled by the controller(a recurrent neural network)and Proximal Policy Optimization(PPO)are two parts,searching for specific data enhancement expansion strategies for specific data sets,greatly improving the diversity of data.Another sub-module is based on Classification model of breast cancer pathology images in ResNet50 network.The experimental results strongly indicate that the diversity of data brought by the AA enhancement strategy effectively improves the performance of the model on multiple evaluation indicators.(3)Breast cancer multi-classification model based on SENet multi-scale hybrid convolutional neural network is proposed.Based on the structural characteristics of the residual block,combined with the SENet attention mechanism and multi-dimensional convolution,a new network module based on the SEET-based multi-scale hybrid convolutional neural network block(SEMSNet)is proposed.And use this module to replace the residual module to propose a multi-scale hybrid convolutional neural network based on SENet.The innovation of this model is to use multi-scale convolution kernel to extract the features of pathological images.From different receptive fields,the pathologists are used to comprehensively distinguish the pathological image properties on pathological images with different magnifications.The model also introduces the SENet attention mechanism module,which is a feature re-calibration attention strategy,which re-calibrates the weight according to the convolution information of each channel,so that the network can acquire more image features.And unlike the traditional residual module,the MSHSNet module uses the overall structure of the inverted residual,and the first"expansion" operation of the inverted residual allows more feature information to be preserved.In the experimental section,three models were trained on the non-enhanced,manual-enhanced,and auto-enhanced data sets.The performance of each model was tested and compared on the test set.The effectiveness of the AA enhancement was also verified on the side.At the same time,in order to prove the validity of the new proposed network model,the ResNet50 training model is used vertically on the automatically enhanced data set and compared,and the results of the new model are compared horizontally with the previous research results.The experimental results strongly indicate that the MSHSNet module structure proposed in this paper has great advantages for feature extraction,which makes the model perform better on multiple indicators.