Research On Pathological Images Analysis And Assisted Discrimination Of Breast Cancer

Breast cancer is one of the major diseases that seriously threaten women's health,and its incidence is getting higher and higher.Histopathological examination based on pathological images is an important basis for the diagnosis of breast cancer,and early diagnosis of breast cancer is the key to improve the success rate of treatment.Because of the characteristics of large dimension,complexity and diversity of pathological images,it leads to high threshold and long time consuming.In addition,due to the subjective differences of different pathologists on the evaluation criteria of tissue and cell characteristics,the consistency between different diagnosis results is low.With the rapid development of digital pathology and computer technology,assisted diagnosis methods based on breast cancer pathological images have become research hotspots.Therefore,according to the actual needs in the diagnosis of breast cancer,this article researches the benign and malignant classification of breast cancer pathological images,the evaluation method of invasive breast cancer cell nuclear atypia,and the evaluation method of invasive breast cancer mitotic figure and related technologies for breast cancer pathological image analysis and lesion-assisted discrimination method.The main work and innovations of the paper are as follows:(1)The classification of benign and malignant breast cancer pathological images was studied.Aiming at the problem that breast cancer pathological images are complicated and changeable,and traditional image processing methods are difficult to effectively extract and combine features,resulting in low classification accuracy,a convolutional neural network-based benign and malignant classification method for breast cancer pathological images is proposed.This method uses convolutional neural network to automatically combine features to realize image classification.Based on transfer learning method,the training weight of convolutional neural network on large natural data sets is used as the initial training weight of the classification model,which solves the problem of model training from scratch.The problem of difficult convergence has proved the feasibility of transferring natural image knowledge to medical pathological images,and achieved a high classification accuracy.(2)The evaluation method of nuclear atypia of invasive breast cancer cells was studied.In view of the existing evaluation methods of nuclear atypia based on region of interest segmentation and artificial features,which are tedious in feature design and low in generality,a fusion strategy-based nuclear atypia evaluation method was proposed.Firstly,a nuclear atypia evaluation method based on convolutional neural network was proposed to directly predict the nuclear atypia score of the target pathological image.Subsequently,a method for scoring nuclear atypia of target pathological images by predicting local labels was proposed.Finally,for the problems of single model predictive tendency and poor stability,the fusion method was used to combine the two methods,and achieved better results on the test set.(3)The evaluation method of mitotic figures of invasive breast cancer was studied.Aiming at the problems of existing mitotic cell detection methods based on statistics,morphology,texture and other artificial features,the feature design is difficult,the small cell detection rate is low,and false positives are high.A mitotic cell detection method based on feature pyramid network and classification verification model is proposed.This method utilizes a feature pyramid network to achieve multi-scale feature fusion for mitotic cell detection,and rejects mitotic cells based on the classification verification model,and achieves higher accuracy and recall on the test set.In addition,in order to solve the problem that it is difficult to determine the effective area of mitotic cells in non-pixel-labeled mitotic cell datasets,a mitotic cell labeling model using existing pixel-by-pixel labeling datasets is proposed.The full text is aimed at the actual needs in the pathological diagnosis of breast cancer.The method of benign and malignant classification of breast cancer pathological images,the evaluation methods of invasive breast cancer cell nuclear atypia and mitotic figure are studied,which can help pathologists in the actual diagnosis and has potential application value.