Research On Pathological Image Classification Of Breast Cancer Based On Multi-view Transformer Coding And Embedded Fusion Mutual Learning

Breast cancer is one of the most common malignant tumors in the world,which has caused great harm to women’s physical and mental health.Computer aided diagnosis system can efficiently assist pathologists in completing pathological image classification of breast cancer.With the rapid development of artificial intelligence,deep learning technology has made significant progress in medical image analysis.However,the existing research has the following shortcomings: on the one hand,the traditional machine learning method based on manual features needs a lot of time and energy,and the feature extraction is insufficient as well as the accuracy is low;On the other hand,the extracted image features based on a single network do not fully mine and utilize the pathological information of breast cancer contained in the heterogeneous networks;Finally,the traditional convolutional neural network lacks the ability to extract global context information from breast cancer pathological images,and ignores the important role of global features in breast cancer image classification.Therefore,this research work focuses on the breast cancer image classification combining multi-view Transformer coding and embedded fusion mutual learning,mainly includes:Breast cancer pathological image classification model based on convolution neural network and Transformer: To solve the problem of insufficient feature extraction of the traditional machine learning method,this study uses many deep learning models,including ResNet50,VGG16,and Transformer,to automatically extract features containing rich pathological information,to accurately depict the lesion areas in breast cancer pathological images,and complete breast cancer pathological image classification.The experiment shows that the ResNet50 model performs best among all the deep learning models.Its classification accuracy on the BreakHis dataset is 89.04%,and the classification accuracy on the BACH dataset is 84.44%.Because only a single network is used to extract features,the key features in breast cancer pathological images are not fully mined,and its overall classification performance needs to be improved.Breast cancer pathological image classification model named EFML based on adaptive feature fusion and embedded fusion mutual learning: To solve the problem that a single network does not fully mine and utilize the key features in breast cancer pathological images,an embedded fusion mutual learning method is proposed: mutual learning between two parallel heterogeneous networks tries to form two effective feature extractors;The adaptive feature fusion classifier and integrated classifier are designed in turn and embedded into the mutual learning training,and the complementary pathological knowledge is learned from the feature map and Logits output.Then the proposed EFML model combines Logits output and fused feature map to complete model training,and realizes pathological image classification of breast cancer.The experimental results show that all performance metrics of the EFML model have been significantly improved,and the classification accuracy of the ResNet50 model in EFML framework on the BreakHis dataset is 99.66%,and the classification accuracy on the BACH dataset is 98.96%.Due to the weak ability of traditional convolutional neural network to capture global context features,the discriminative ability of EFML needs to be improved.Breast cancer Pathological Image Classification Model called MVT-EFML Based on Multi-view Transformer Coding and Embedded Fusion Mutual Learning: The traditional convolutional neural network lacks the ability to extract the global context features in the images.Mixup data enhancement algorithm is first used to generate high-quality pathological images of breast cancer;Then a dual-stream network structure combining ResNet50 and Transformer is designed: ResNet50 is responsible for extracting local features in the pathological images;The multi-view Transformer coding module is responsible for encoding the spatial position relationship of multi-view images to better capture the global context features in the pathological images;The model training is completed based on the embedded fusion and mutual learning methods,and the classification of breast pathological images is realized.The experimental results show that MVT-EFML achieves 99.77%accuracy on the BreakHis dataset,and the mean value of F1 index is 4.75% higher than the state-of-the-art baseline.The accuracy reaches 98.97% on the BACH dataset,and the F1 metric is 3.21% higher than the best baseline.The ability to locate local and global features of pathological images of breast cancer is enhanced,and the interpretability is further improved.