Molecular Subtype Classification Of Breast Cancer Based On Deep Learning

Breast cancer is a malignant tumor that is common in women.Early detection of breast cancer and targeted treatment of different breast cancer subtypes can greatly reduce patient mortality and reduce patient suffering.Doctors analyze the patient’s imaging report to confirm the patient’s condition,which is often time-consuming,laborious,and inefficient.Therefore,the construction of computer-aided diagnosis system for breast cancer mammogram images has great application value and practical significance.At present,the research methods on molecular subtype classification of breast cancer are mainly divided into traditional machine learning methods based on manual feature selection and classification methods based on deep learning.Traditional classification methods rely on prior knowledge to a greater extent because of their artificially selected features,resulting in poor generalization.The classification method based on deep learning is an end-to-end model,which can avoid the incompleteness and bias of artificial selection of features,and the recognition accuracy is greatly improved compared with traditional image classification methods.However,deep learning-based methods generally require large-scale datasets for training and clinically experienced physicians to label lesions.Therefore,building models suitable for small-scale datasets based on deep learning becomes a huge challenge.In view of this,based on deep learning,this paper studies a small number of breast cancer mammogram images,and divides breast cancer mammogram images into two categories: Luminal type and FLuminal type.(1)A dual-input ESE-Mobile Net model based on lightweight neural network Mobili Net-v3 and dual-view strategy is proposed.First,for a small amount of breast cancer data,we propose to use a lightweight convolutional neural network Mobili Net-v3 to reduce the amount of network training parameters and computational complexity,effectively reduce the risk of network overfitting,and add an attention mechanism ESE module to the Bottle Neck module of Mobili Net-v3 to improve the feature extraction ability of the network.Second,in view of the characteristics of two views(head and tail position,inner and outer oblique position)of unilateral breast mammogram examination,we propose a dual-view classification strategy to integrate the information of the dual-view image of unilateral breast,so as to make up for the limitations of the existing single-view method,which is in line with the diagnostic process of radiologists.The designed dual-input ESE-Mobile Net model performs well in various classification evaluation indicators.Through a series of sufficient comparative experiments,the effectiveness of the dual-input ESE-Mobile Net model constructed in this paper is proved.(2)Another lightweight neural network Reg Net is proposed,the channel space attention ESAM module is introduced,and the improved algorithm model is called ESAM-Reg Net,and two hybrid models are constructed based on two ensemble strategies.Specifically,firstly,the ESAM channel spatial attention module is introduced in the feature extraction layer of the Reg Net model,which is used to improve the model’s attention to important feature information and ignore irrelevant feature information.In order to avoid the shortcomings of a single classifier and achieve better classification effect,we fuse the two lightweight networks we built ESE-Mobile Net and ESAM-Reg Net based on the two integration strategies of weighted average method and stacking method.Experimental results show that the comprehensive classification performance of the hybrid model is improved.Compared with other ensemble models and mainstream CNN networks,the effectiveness of the algorithm improvement is proved.