Breast Cancer Histopathological Image Classification Based On Deep Learning

Breast cancer is the most common cancer in women worldwide.Early diagnosis is essential for the treatment of breast cancer.Despite the rapid development of medical technology,histopathological diagnosis is still considered the best way to diagnose breast cancer.However,the complexity of histopathological images and the sharp increase in workload make this task very time-consuming.At the same time,the knowledge reserve of pathologists will affect the accuracy of diagnostic results.Therefore,the development of automated and accurate breast cancer histopathological image classification system is of great clinical significance.This topic is based on deep learning to classify and predict breast cancer histopathological images.The main research contents of the subject are as follows:(1)A breast cancer histopathological image prediction method based on transfer learning and attention mechanism is proposed,which is called dual information flow mobile inverted bottleneck convolutional neural network(DMBC-Net).Firstly,a new breast cancer image classification network,DMBC-Net,is constructed.The DMBC-Net contains multiple basic DMBC modules,which fuse coordinate attention and depth separable convolution.By extracting channel information,the feature map with more effective information is obtained while expanding the receptive field of the convolutional neural network.Secondly,DMBC-Net is trained for secondary transfer learning on the basis of Image Net and BACH datasets,and the model parameters are optimized by transfer learning.Finally,the DMBC-Net model is verified on the public Brea KHis dataset,and four evaluation indicators are used to evaluate the performance of the independent model.The experimental results show that DMBC-Net has excellent prediction ability in the binary classification task based on amplification correlation,and solves the problem of high false positive of most prediction models.Comparing DMBC-Net with other advanced models,DMBC-Net is superior to common advanced algorithms.(2)A breast cancer histopathological image recognition algorithm based on pure attention structure is proposed,which is called SAL-Vit model.Firstly,SKIPAT parameter function is constructed by SKIPAT method,and SKIPAT parameter function is used to replace multiple multi-head self-attention modules in Vision Transformer(Vi T)model to construct SA-VIT model.The model retains the excellent performance of Vi T while reducing the dependence on datasets,maintaining fewer parameters and lower computational complexity.Secondly,the SL-Vit model is constructed.By shifting the patch labeling method,the space is added to the visual marker in the image slicing step,and then the diagonal shielding and learnable temperature scaling are used to enhance local self-attention.Local self-attention can enhance the bias induction ability of SL-Vit model.Finally,the integrated model SAL-Vit is constructed by soft voting integration strategy,and SAL-Vit is verified on Brea KHis dataset and BACH dataset.At the same time,a series of ablation experiments are designed to verify the effectiveness of each method and integration strategy.The experimental results show that the SAL-Vi T model can classify the pathological images of ui breast cancer more accurately,alleviate the dependence of the Vi T model on the huge data set,reduce the computational complexity of the model,and train faster.According to the results of the SAL-Vit model,it can provide objective results support in the doctor ’s diagnosis process,and provide the doctor with a second reference suggestion to assist the doctor in the diagnosis of breast cancer.