The Research Of Colon Cancer Histopathological Image Classification Based On Deep Domain Adaptation

With the changes in people’s dietary habits,the global incidence and mortality of colon cancer are increasing.Now colon cancer has become one of the main diseases threatening people’s health.Clinically,the diagnosis method based on histopathological images of colon is the standard method for detecting colon cancer.Deep learning method in computer-aided diagnosis has been widely used in the histopathological diagnosis of colon cancer in recent years.Deep learning relies heavily on samples and requires a large amount of labeled data to understand the potential information.However,the collection of colon cancer histopathological images is complicated and expensive.It is very difficult to establish a large-scale labeled colon cancer histopathological images dataset.Transfer learning can alleviate the problem of insufficient labels to a certain extent.However,the models based on traditional transfer learning methods are pre-trained on natural images,and the adaptability of the models for the classification task of colon cancer histopathological images is unsatisfactory.The performance of the models remains to be further improved.In order to address the above problems,based on the characteristics of different colon cancer histopathological images,this thesis studies and designs two transferred semi-supervised domain adaptation classification and recognition algorithms.To solve the problem of insufficient labels,deep transfer learning is used to extract the high-level features of the dataset.Meanwhile,the semi-supervised domain adaptation algorithm is used to improve the classification adaptability of the model.The main research work of this thesis includes the following contents:A semi-supervised domain adaptation classification algorithm for histopathological images of Minmice and patients based on deep transfer learning is studied.Firstly,the deep information of the images and the high-level feature representation for texture and color of the images is extracted through the transferred neural network.Then,a dual-criterion semi-supervised domain adaptation algorithm is used to align the high-level features of the Minmice and patients images to narrow the feature differences between the two domains.The conditional entropy term which can reveal the underlying characteristics of the patients’ samples in the target domain is studied.It promotes the perception of the data information in the target domain and enhances the classification performance of the model for the patients’ dataset.A transferred semi-supervised domain adaptation classification algorithm for Whole Slide Images(WSIs)of patients with colon cancer based on a multiple weighted loss function is studied.Firstly,a convolutional network is designed to extract high-level features of histopathological images.Then,a semi-supervised domain adaptation algorithm based on a multiple weighted loss function is used to fully align the high-level feature distributions of the patients’ datasets in the source domain and the target domain.Manifold regularization term is used to improve the adaptability of the model to the data distribution space of the target domain and strengthen the classification and recognition performance of the model.The research work in this thesis solves the problem of insufficient label information in the dataset in the classification of colon cancer histopathological images.Deep transfer learning alleviates the need of the model for large amounts of labeled data.The domain adaptation algorithm improves the classification performance.This research provides a feasible solution for the computer-aided clinical diagnosis of colon cancer and has certain theoretical value and reference significance.