Research On Breast Cancer Pathological Image Recognition Model Based On Multi-angle Deep Feature Mining

Breast cancer is one of the most common cancers in women,which poses a serious threat to women’s physical and mental health.The computer-aided diagnosis model can accurately and efficiently assist the clinical diagnosis of pathologists,realize the rapid diagnosis of breast cancer,and lay an important foundation for improving the quality of life of patients.However,the existing research still has the following deficiencies: on the one hand,only a single model is used to mine pathological information,and the feature complementarity between different models cannot be used;on the other hand,the global features of highresolution pathological images are ignored in breast pathological images The importance of recognition,and the diagnostic results lack interpretability;in addition,the traditional Transformer cannot make full use of the spatial feature information in high-resolution pathological images.To this end,this study conducts deep feature mining on breast pathological images from multiple perspectives such as convolutional neural network,multiinstance learning,and Transformer,and proposes a breast cancer pathological image recognition model based on multi-angle deep feature mining.The main work is as follows:(1)Model Fusion based on Online Mutual Knowledge Transfer: Aiming at the problems of insufficient extraction of pathological features and limited ability of recognition and generalization for a single model.Starting from the real pathological diagnosis scenario,model fusion based on online mutual knowledge transfer(MF-OMKT)is proposed.MFOMKT realizes online mutual knowledge transfer through deep mutual learning,and strives to break the isolation between models,which helps to complete deep feature mining and provides complementary information from heterogeneous networks for model fusion.MFOMKT also designed an adaptive model fusion method to fuse the heterogeneous network features after knowledge transfer,and train a model with stronger recognition performance and generalization ability.On the Break His dataset,the accuracy rate of MF-OMKT for binary classification is over 99%,and the accuracy rate for multi-classification is over 96%.The MF-OMKT model is closer to the real pathological diagnosis scene: online mutual knowledge transfer simulates the mutual communication and learning between pathologists,model fusion simulates the process of pathologists obtaining diagnostic results through centralized decision-making,and MF-OMKT is a general-purpose model fusion framework.(2)Multi-View Attention-Guided Multiple Instance Detection Network: MFOMKT takes image patches in pathological images as model input and achieves better performance in small and medium resolution images,but it is not suitable for high resolution pathological images.To deal with this problem,a multi-view attention-guided multiple instance detection network(MA-MIDN)is proposed.MA-MIDN transforms the traditional image classification problem into a weakly supervised multi-instance learning problem,and designs a multi-view attention algorithm to screen examples to locate key regions in the image.And a multi-view attention-guided multi-instance pooling method is designed to aggregate instance features to obtain more discriminative image-level pathological features.The MA-MIDN model simultaneously completes lesion localization and image classification.The accuracy,AUC,Precision,Recall and F1 indicators of the MA-MIDN model on three public breast cancer pathological image datasets are all better than the baseline.It achieves precise positioning without affecting classification performance,provides more interpretable diagnostic results,and has high practicality.(3)Deep Hybrid Vision Transformer: MA-MIDN performs well in the binary classification task of high-resolution pathological images,but its multi-classification performance needs to be improved,and the spatial information in the image cannot be effectively utilized.To this end,a recognition model based on deep hybrid vision transformer(DH-Vi T)is proposed.DH-Vi T combines Transformer and Convolutional Neural Network,and adds a pyramidal convolution structure to the Transformer structure to fully extract the spatial information in high-resolution pathological images.In addition,the online knowledge distillation is introduced into the DH-Vi T model to deeply mine the pathological information in the pathological images of breast cancer and improve the accuracy of the model.On three public breast cancer pathological image datasets,the DH-Vi T model outperforms mainstream baselines and significantly improves performance on multi-classification tasks.