Research On Multi-Classification Of Breast Histopathological Patterns Based On Deep Learning

Breast cancer is one of the types of cancers with high incidence and mortality worldwide,which seriously endangers the health of female groups.Histopathological examination is the "gold standard" for cancer diagnosis.However,the inherent complexity and diversity of breast histopathology images make the diagnosis of pathologists non-trivial and time-consuming.In addition,the differences in experience between pathologists and the subjectivity of pathological diagnostic criteria often lead to inconsistency and non-reproducibility of diagnostic results.Computer-aided automatic diagnosis of breast cancer histopathology images can help pathologists improve the consistency and efficiency of diagnosis,and reduce the intensity of work.With the development of deep learning,the convolutional neural networks have gradually replaced the hand-crafted feature extraction methods designed for specific problems based on domain knowledge,and automatically extract the features of images for classification,which are widely used in the field of medical pathology analysis.Above all,based on deep learning and hematoxylin-eosin stained breast histopathology images,computer-aided multi-classification methods in the field of histopathology were studied.The main research contents of this thesis are as follows:(1)This thesis proposes a model based on sampled patches processing to classify high-resolution breast histopathology images into four categories.Firstly,based on the atypia and disarrangement of the nuclei,and the differences between morphological structure of the normal tissue and the diseased tissue,two different sizes of patches are sampled from the histopathology images to contain nucleus-level and tissue-level features.Secondly,for the lack of diagnostic information or label errors of some patches,a batch screening method based on a convolutional neural network and clustering algorithm is proposed for selecting more discriminative patches.Finally,feature extraction and fusion are performed on the sampled patches to obtain image-level feature for classification.The model achieves 95%accuracy on the initial test set released by the 2015 Bioimaging Competition.The comparison experiments also verify the effectiveness of the two strategies.(2)A two-stage classification framework is proposed to classify the subtypes of benign and malignant lesions for patients based on the dataset of breast histopathology images at four magnification rates organized in patients.In the first stage,a model is proposed to classify the patients into benign or malignant.Firstly,the convolutional neural networks are used as classifiers to classify the breast histopathology images acquired by each patient at different magnification,and then the meta-decision tree is used to combine the predicted results of four classifiers.In the second stage,the proposed model is used to classify the sub-types of the benign and malignant lesions.The convolutional neural networks are used as the feature extractors,and then random forests are constructed based on the image feature groups obtained by each patient under different magnifications.According to the category information,the corresponding dissimilarity matrix is obtained,and fused by averaging to compute the joint similarity matrix as the kernel of the support vector machine.The framework achieves 92.5%accuracy of 8 categories on BreaKHis,which proves its effectiveness in multi-classification of breast pathologic sub-types.