Study On Breast Cancer Classification Method Based On Pathological Images

Breast cancer is one of the leading causes of death from cancer in women.The incidence rate shows an increasing trend and the age of onset is with a younger trend.Diagnosis based on pathological images is considered as the golden standard in the clinic.At present,the classification of breast cancer pathological images has the following problems:(1)The acquisition of pathological images requires a series of complicated processes such as tissue preparation,scanning,and labeling,which is costly.Therefore,the clear and well-marked breast cancer pathological image datasets are few.(2)The diagnosis results are often based on the subjective judgment of the pathologists,and the workload is large.Moreover,the need for accurate diagnosis and treatment is increasing.(3)Breast cancer pathological images have complex morphological structure,uneven staining,adherent nuclei and other conditions,making it more difficult to identify breast cancer.Computer aided diagnosis system is helpful to reduce diagnosis cost and improve diagnosis efficiency and objectivity.Based on digital pathological imaging technology,pathological image classification methods are generally divided into classification based on artificial feature extraction and classification based on deep learning.Based on the western databases,some key problems in the above two classification methods of breast cancer pathological images were deeply studied in this thesis.The main work and innovations of the thesis were summarized as follows:(1)According to the problems of sparse chromatin,uneven staining and adherent nuclei in the nuclei segmentation of breast cancer pathological images,a 3-output convolutional neural network(CNN)segmentation method was designed.By replacing large convolution kernel with multiple small convolution kernels,the number of parameters and the computational complexity were reduced.Probability growth method was used to post-process the probability maps obtained from CNN model to further improve the accuracy of nuclei segmentation.The experimental results showed that the method was more effective for nuclei segmentation than the existing methods.(2)In view of the different classification and discrimination ability of different color spaces in texture feature extraction,combined with the characteristics of image acquisition in multiple color spaces,a texture feature extraction method based on H&E(Hematoxylin&Eosin)model was proposed.For the rich geometry and complex texture caused by the morphological and structural diversity of benign and malignant breast cancer nuclei,morphological,spatial and texture features were integrated to express the biological characteristics of breast cancer image,which achieved the improvement of classification performance of benign and malignant breast cancer.(3)For the small dataset of breast cancer pathological images,combined with the knowledge transfer characteristics of transfer learning and the powerful ability of feature representation of deep network model,the classification method of CNN model as feature extractor and fine-tuning CNN model was compared and analyzed.The experimental results showed that fine.tuning CNN model was more effective for the classification of benign and malignant breast cancer in the database without segmentation annotation.