Classification Of Breast Cancer Pathological Images Based On Feature Map Visualization And Feature Fusion

Breast cancer is a common disease among women,and its clinical diagnosis needs to be based on the analysis of breast histopathologic images.Due to the complexity of breast tissue structure and cell morphology,the diagnosis process is tedious,and requires extremely high professional level of pathologists.With the continuous development of deep learning technology in the field of image recognition,classification models based on such algorithms can automatically extract discriminant information from input images,independent of cell segmentation and traditional artificial feature extraction,showing great advantages.In this thesis,the deep learning method is used to classify the pathological images of breast,and a network optimization strategy based on feature visualization and a high-precision automatic classification approach based on deep feature fusion are proposed.The main works are as follows:Firstly,in order to solve the problems of unnecessary computing resource loss and overfitting caused by the large amount of traditional network training parameters,a network parameter optimization strategy based on feature map visualization is designed,and the VGG16 network structure is improved for breast cancer pathological image classification.The feature maps generated after convolution operation of input images is output by the visual model,and the feature information extracted from the network is observed in a more intuitive way,and the network depth and the number of convolution kernels are determined by combining the information entropy.The network parameters are reduced on the premise of ensuring the model performance.Secondly,a data enhancement method is designed for breast histopathologic images.This method uses image transformation,image patch generation and staining normalization techniques to preprocess the images in the training set,aiming at solving the problem of insufficient samples in the data set caused by the difficulty of patient data collection,and alleviating the influence of too deep or too shallow coloring on the image quality during the staining process of pathological slices.Furthermore,a high precision automatic diagnosis approach for breast histopathologic images based on feature fusion is proposed.To solve the issue that the image classification model usually only focuses on the single feature dimension of the input image,and cannot completely represent the details of cell shape,contour and texture in pathological images,the cascaded feature fusion strategy and attention mechanism are introduced into the pathological image classification method,and the multi-scale feature fusion strategy is designed.By emphasizing the important feature information mapped in the input image in the network,the model can improve the discrimination of the structural differences between benign and malignant cells,and realize the cascaded integration of the depth feature information of different convolutional layers,so that the features learned by the network can realize the integration of local and global as well as spatial and channel.In this way,more expressive depth features can be obtained.Finally,by comparing the existing methods based on single network and hybrid network on the public data set,the classification approach of breast tissue images proposed in this thesis based on feature map visualization and feature fusion can realize the effective diagnosis of benign and malignant breast tumors on different indicators,which provides a certain theoretical basis for computer-aided diagnosis of breast cancer.