Classification Of Breast Cancer Histopathological Images Based On Convolutional Neural Networks

The model of convolutional neural networks(CNN)has a very excellent learning ability.It inputs image data into the network directly,and processes the image with its very careful feature extraction method without the complex operations of preprocessing and extra feature extraction.Compared with the traditional machine learning algorithm,CNN has achieved good results in research fields such as image classification,image segmentation,and target detection.In recent years,CNNs have been applied to the medical image analysis because it can automatically learn and extract the underlying diagnostic features of disease from medical image.However,the histopathological images of breast cancer are relatively complicated and have wide heterogeneity.The current methods of automatic image classification have limitations in high-resolution image processing,which makes the final classification effect less ideal.Therefore,accurate evaluation of cell morphological features is still a major problem.To solve the above problems,this paper designs two kinds of breast cancer histopathological image classification models based on the CNN.The research contents are as follows:(1)To make the network performance more perfect and improve the feature extraction ability of the model,the current CNN is developing in a direction of deeper layers.However,the process will increase the number of model parameters,make the training process more complicated,and easily lead to the risk of over-fitting.To solve the above problem,a new small Squeeze-and-Excitation Residual Networks(SSER)is designed in this paper.Compared with the existing Squeeze-and-Excitation-Residual Networks,the SSER module can obtain the competitive classification performance with fewer parameters.Experimenting on the Cifar natural image dataset,we found that compared with the bottleneck SE-ResNet module and the basic SE-ResNet module,the SSER module has reduced the number of parameters by 29.4% and 33.3%,respectively,and can obtain a similar classification results;(2)In the process of network training,it is usually necessary to manually and continuously adjust parameters.To solve this problem,a new learning rate scheduler named the Gaussian error scheduler(ERF)is designed in this paper.Compared with other learning rate schedulers,the ERF is relatively simple in the process of network training and the classification performance of trained network model is better;(3)In this paper,we designed two models for the classification of breast cancer histopathology images.The first model named the breast cancer histopathology image classification networks(BHCNet),which is composed of SSER module,pooling layer and the fully connected layers.Compared with the existing network,the BHCNet can effectively avoid the complexity and limitation of manual feature extraction under the condition of fewer parameters and higher classification accuracy.In this paper,BHCNet is used for the automatic classification task of breast cancer histopathological images.The accuracy of binary classifications is about 98.87% to 99.34%,and the accuracy of eight classifications is about 90.66% to 93.81%.We further applied the BHCNet to the small breast histopathology images,and found that its performance was normal through experiments.Therefore,a transfer learning structure was introduced.A convolutional neural networks model based on transfer learning is designed to classify the small breast cancer histopathological image datasets using pre-trained models that have been fully trained on large datasets.The accuracy of the binary classification(no cancer,cancer)was 91.67%,and the accuracy of the four classification was 86.11%.