Research On Weakly Supervised Fine-grained Breast Cancer Pathology Classification

In the new century,the continuous development of science and technology and the continuous progress of social production have brought earth-shaking changes to people’s lives,but also brought some problems,such as environmental pollution,accelerated pace of life,increased work pressure,and so on,which further lead to health problems.Cancer,one of the diseases with the highest mortality rate,has a higher prevalence in recent years,and shows an obvious trend in younger people.Breast cancer,as a major"killer" of women’s health,deserves our attention.For breast cancer patients,early diagnosis and early treatment can significantly improve the survival rate and prognostic quality of life.The diagnosis of breast cancer can be based on a variety of imaging images,but the pathological image is considered the gold standard for cancer diagnosis.However,The way of manual classification by pathologists may be inaccurate due to the differences in the process of pathological image preparation,the complex changes of tissue structure and the subjective understanding of pathologists.With the development of artificial intelligence technology,computer-aided medical diagnosis emerges at the right moment,and has a good application in the field of breast cancer pathological image classification.Although the existing methods of breast cancer pathological image classification have achieved satisfying results,there are still some problems to be solved in this field.1)Computer-aided medical methods need good interpretability,while most of the existing classification methods of breast cancer pathological images only focus on the improvement of accuracy,but ignore the performance of model interpretability.2)The existing pathological image classification of breast cancer is mostly simple benign/malignant binary classification,which has a limited auxiliary role in clinical diagnosis.Further fine-grained classification is more clinically significant.3)Pathological images are very different from natural images.In pathological images,both background and targets belong to tissue regions without clear boundaries,and there are multiple scattered targets in one pathological image.The existing pathological image classification methods mostly consider the problem from the angle of natural image classification,without considering the particularity of medical images.4)The medical image classification model needs high accuracy,and the existing methods mostly improve the accuracy by refining the model structure.As a result,the model structure is complex and it is difficult to achieve the balance between high accuracy and simple structure.In this thesis,the author conducted a study on weakly supervised fine-grained breast cancer pathological image classification methods,and designed two weakly supervised fine-grained breast cancer classification methods to solve the above problems.The main contributions of this thesis are as follows:(1)To address the limitations of existing binary classification tasks,this thesis proposes two fine-grained classification methods,which can accurately distinguish each subtype of breast cancer and provide more clinical reference information.(2)Aiming at the problems of insufficient interpretability of existing models and failure to consider the particularity of medical images,the author proposed a weakly supervised fine-grained method for breast cancer classification based on deep features.The method uses convolutional neural network and attention network to simulate the actual diagnosis process of doctors,and has good interpretability.Image cropping and image dropping are also introduced to enable the model to focus on multiple suspicious regions.In addition,considering the particularity of medical images,a hierarchical cross-entropy loss is designed by using the hierarchical label of medical images to improve the classification accuracy.(3)Aiming at the problem that the existing methods are difficult to achieve the balance between high accuracy and simple model structure,the author proposed a weakly supervised fine-grained method for breast cancer classification based on sequential features.Based on Vision Transformer(ViT)model,this method is the first to apply ViT to fine-grained classification of breast cancer.In addition,the method improves the original ViT model based on the characteristics of medical image fine-grained classification to make it more compatible with the characteristics of medical image fine-grained classification,and the inherited ViT model’s self-attentive mechanism simplifies the model structure while achieving a high accuracy rate,achieving a balance between high accuracy and simple model structure.(4)Extensive experiments on three commonly used pathological image datasets of breast cancer demonstrated the effectiveness of the two fine-grained classification methods proposed in this thesis.