Pathological Image Classification Of Breast Cancer Based On Deep Learning

Pathological images contain a wealth of phenotypic information that can be used to monitor patient condition and prognostic treatment.Pathologists through examination and interpretation of stained pathological sections to diagnose and assess early disease.However,this manual diagnosis is inefficient,has poor classification accuracy and the workload of doctors is heavy.At the same time,the emergence of computer-aided diagnostic systems has reduced the pressure on pathologists to read the slides and reduced the rate of misdiagnosis.In the past few years,with the development of deep learning in the field of pathologiy-aided diagnosis,feature extraction methods based on deep learning have been widely applied to the classification of breast cancer pathology images.However,due to the complex and diverse morphological structures of breast cancer images,some existing deep learning methods do not fully utilize the tissue and texture features.Therefore,in order to further improve the classification accuracy of breast cancer pathology images,this paper investigates the problem of classification of breast cancer pathology images from the perspective of deep learning feature extraction and feature fusion enhancement.The specific research contents are as follows:(1)Aiming at the problem of insufficient feature extraction of breast cancer pathological images in the task of weakly supervised classification,a pathological images classification method using unsupervised feature extraction is proposed.In this method,small patched without label information are first cut out from the original pathological whole slide image.Then this method uses the unsupervised learning method BYOL to target training on the feature extraction network so that the feature extraction model can mine more adequate pathology tissue features.Finally,features are attention pooled and classified using a multi-instance weakly supervised framework with clustering constraints.Illustrated by experiments,the proposed method can effectively classify the panoramic images of breast cancer pathology and improve the classification accuracy of the network using only image-level weak labels.(2)Aiming at the problem of multiple magnification dimension feature is not comprehensively used in the task of breast cancer pathological image classification,on the basis of weakly supervised classification,a multi-instance weakly supervised pathological WSI classification method that incorporates two-scale feature information is proposed,and named as Multi-CLAMBR.In this method,small tiles of multiple scales are first extracted from the pathological tissue region of the whole slide images following the 1:1 center extraction.Then,the feature extraction model trained by unsupervised contrastive learning is used to extract small patch features of multiple magnifications.Finally,features extractes from multiple magnifications of the patches are fused together to classify breast cancer pathological images.Illustrated by experiments,Multi-CLAMBR further improves the classification accuracy of the network by fusing multiple magnification-scale features with only image-level weak labels,and the classification performance is better than most other two-scale methods.(3)Aiming at the problems of unevenly distributed of data in different categories,small data samples size,and insufficient use of texture information in images in multiclassification tasks of pathological images,a multiclassification method using texture enhancement for breast cancer images is proposed.In this method,first of all,a data augmentation method is used to enrich and balance the data of each category,and then the cell and tissue texture features are characterized by using stain separation technique and Gray Level Co-occurrence Matrix algorithm.Next the extracted feature images are fed into the spatial texture enhancement block,allowing the network further focus on the texture features in the image.Finally,the classifier fuses all features and performs classification.Illustrated by experiments,the texture enhancement-based multi-classification method can significantly improve the multiclassification performance of the original network.In summary,this paper proposes specific solutions to the problems of insufficient feature extraction,insufficient utilization of multiple magnifications feature information in weakly supervised classification tasks,and unbalanced data distribution and insufficient utilization of texture features in multi-classification tasks,starting from the complexity and diversity of breast cancer pathology images.The experiments show that the proposed methods in this paper are effective in improving the classification performance of existing methods.