Subtype Classification Of Breast Tumor Pathological Images Based On Multi-Scale Feature Fusion

In recent years,the prevalence of breast cancer has been rising,gradually surpassing lung cancer as the largest cancer threatening women’s health.Early diagnosis and early treatment can effectively improve the survival rate of patients.Pathological examination,as the "golden criteria" for the diagnosis of breast cancer,is the only basis for determining whether there is a pathological change in its tissue area.Accurate classification of pathological Images of breast tumors through computer-aided diagnosis and treatment systems can help doctors better and faster analyze tissue lesions,which has important application value in the field of clinical medicine.However,the uneven staining of breast tissue Images,the uneven distribution of lesion types,and the complexity and diversity of pathological Image structures make it difficult to continue to improve the classification accuracy.In this paper,a classification method of breast tumor pathological Image subtypes based on multi-scale feature fusion.Combined with the type distribution rule of breast tumor pathological data set,Image structure,and dyeing characteristics,the samples are preprocessed pertinently,and a deep neural network model is built to achieve accurate classification of breast tumor pathological Images.First,an improved pooled residual module is added to Res Net50 network model to suppress the additional noise introduced by random sampling and reduce the loss of breast tissue depth features.Then,a feature fusion unit is designed to splice and reuse the shallow and deep multi-scale features extracted from the network and share them with the classification branch network to improve the feature utilization,while providing necessary high-resolution information for the classification model.Finally,a complementary classifier is established to enhance the internal relationship between the prediction results of benign and malignant tumors and the prediction results of their subtypes through the complementary loss function,and the automatic classification of breast tumor subtypes is realized on the fusion feature by simulating the real diagnosis scene.The comparison experiment on the public data set Break His shows that the average accuracy rate of the proposed method in the four magnifications is 98.98%,the average precision rate is 98.74%,the average recall rate is 98.51%,and the average F1 score is 98.58%.Compared with the comparison method,it not only obtains better classification performance,but also the proposed complementary classifier,feature fusion unit,and pooled residual module can further base the network classification performance,which has a certain clinical value.