Research On Classification Of Breast Histopathological Images Based On Deep Learning

As a disease with extremely high mortality and morbidity,breast cancer can significantly increase the possibility of genetic mutation through multiple carcinogenic factors,which can lead to uncontrolled cell proliferation.The disease is also known as the "pink killer" because its target is mostly women.Histopathological examination is one of the most effective ways to diagnose breast cancer.Although there are many methods of computer-aided cancer diagnosis,most of them cannot accurately diagnose the sub-type of the cancer.Because the pathological images of breast tissue contain complex image information,it is cumbersome and time-consuming for doctors to make accurate judgments,and the diagnostic experience of doctor is different,making the consistency of the diagnostic results low.With the advancement of computer software and hardware,computer-assisted diagnosis of cancer based on pathological images has become a trend.Deep learning methods such as convolutional neural networks have powerful and automatic feature extraction capabilities,therefore algorithms based on convolutional neural networks have been broadly applied in breast cancer classification based on breast tissue pathological image.However,some current breast cancer classification algorithms still have the following drawbacks: they do not pay attention to the multi-scale information in the images;pathological images have noises and complex image information;there are large inter-class difference and small intra-class difference;it is difficult to distinguish the cancer subtype.The change of scale will disturb the stability and discrimination of the feature representation of the pathological images,thus affect the performance of benign and malignant tumor classification.To address these issues,this thesis studied on the classification of breast tomor based on breast histopathological images utilizing convolutional neural network as well as considering the image scale changes.The main contribution of the thesis is described as following:(1)This thesis proposes a benign and malignant classification model of breast histopathological images based on single-scale feature cross.This model includes two modules of semantic grouping and feature enhancement,which are designed to suppress background noise,highlight key features,and improve the discrimination of single-scale inter-class features in breast histopathological images.To verify the effectiveness of the proposed single-scale feature cross model,this thesis conducts binary classification experiments on the Break His dataset.The experimental results show that the feature cross operation for single-scale features can effectively suppress the background noise of the image,enhance the discrimination of key semantic features in the image,and then improve the classification accuracy.(2)In this thesis,a dual-branch network model based on multi-scale feature representation is proposed for multi-classification of breast tumor based on histopathological images.In order to improve the multi-scale feature representation of breast histopathological images,a dual-branch network is designed to extract corresponding global features and key local features,where global features have complete contextual information of pathological images,and key local features have significant core features and small changes in target scale.The key local features are used as supplements to perform multi-scale analysis of pathological images.A fine-grained region localization method is employed to achieve quantification and localization of key local regions.To verify the effectiveness of the model,multi-class ablation experiments and comparative experiments are also performed on the Break His dataset.The experimental results show that the classification accuracy of the dual-branch network model based on multi-scale feature representation is higher.It can be suggested that the multi-scale problem is one of the bottlenecks limiting the performance of medical pathology image classification,and a reasonable multi-scale feature representation strategy can effectively improve the classification performance.