Research On HER2 Classification Method Based On Wide Network

Human epidermal growth factor receptor 2(HER2)gene amplification or high protein expression exists in breast cancer patients,which is called HER2 positive breast cancer.The expression level of HER2 can reflect the disease degree of breast cancer patients,and people have stipulated different scores according to its expression level.The correct classification of HER2 score is related to the accurate use of targeted drugs for breast cancer,and is very important for the prognosis and treatment of breast cancer.Therefore,this thesis studies how to use neural network framework to improve HER2 recognition performance.Determination of the HER2 category relies on the pathologist’s visual observation,which is time-consuming and impersonal.Due to the development of digital pathological images and artificial intelligence,many computer-aided methods for HER2 classification have emerged,but their performance needs to be improved.Based on the above problems,this thesis mainly studies the design of a neural network model for HER2 classification under the premise of reducing human workload and improving performance.Firstly,the features displayed after dyeing HER2 image film are susceptible to interference from other factors,so targeted image preprocessing is needed.In this thesis,a color domain conversion method is proposed to preprocess the HER2 image by limiting the range of HSV components.This method first converts RGB color domain to HSV color domain,then screens the required brownish and yellow color features,and finally returns to the original RGB color domain.Experimental results show that this method highlights the membrane staining characteristics of HER2 and removes other interference factors such as color and light.Secondly,this thesis compares a large number of classification neural network models through experiments,and finds that Google Net has a good effect on the classification of HER2.Therefore,in view of the advantage of Google Net model’s "width" for HER2 classification,this thesis proposes a width-expanding neural network model WPNet.By using convolution kernels of different sizes in parallel,this network model can provide different receptive fields and effectively extract film dyeing features of different sizes in HER2 images,thus improving the classification effect of HER2.Through a large number of experiments,this thesis proves that WPNet has a higher accuracy for HER2 classification than other models.Finally,because the input resolution of WPNet model is larger and the convolution kernel of larger size is used in the model,the calculation cost of WPNET model is larger than that of conventional lightweight networks.Therefore,this thesis proposes a lightweight network model LWPNet suitable for HER2 classification on the basis of width network.The combination of packet convolution,expansion convolution,channel mixing and width network not only reduces the calculation cost,but also ensures the classification accuracy of HER2.Experimental results show that compared with other models,LWPNet has lower calculation cost and higher classification accuracy.