Pathological Image Analysis Based On Deep Learning

Histopathological images are the gold standard for cancer diagnosis.The patholo-gists explore the causes,pathogenesis,and development process of the lesions to make a pathological diagnosis by analyzing the pathological images.However,on the one hand this work is time-consuming and labor-intensive,and on the other hand it relies heavily on the subjective judgment and operation of the pathologists.Therefore,the development of computer-aided diagnosis has high practical significance.The analysis of pathological images can be carried out from two levels of tissues and cells.So how to conduct intelligent analysis on two levels is particularly important.This research focuses on pathological image analysis and the main work is as follows:1)A Branch-Aggregation Network based on deep learning is proposed for patho-logical image classification.First,the lung pathological images obtained from the hos-pital are processed to construct a lung pathological image dataset.Then the proposed network is used for training and testing under low magnification images.Comparison of the experimental results of the proposed method with the variant models and the com-petitive models proves the effectiveness of the proposed network.Finally,the patches of pathological slices are classified under the images of high magnification,and the classification results are used to realize the preliminary segmentation of the tumor area.2)A weakly supervised method based on W-shaped segmentation colorization net-work is proposed for nuclei segmentation by point annotations.First,the hematoxylin component extracted from the H&E stained pathological image is as the model input,and two kinds of coarse labels as the supervision information are obtained by process-ing the point annotations.Then the proposed network is trained by co-training to learn more features,and colorization is introduced as an auxiliary task to improve the learning ability of the network.Finally,the comparison results on the public dataset with other weakly supervised segmentation algorithms show that the proposed method is superior to competitive algorithms and has practicality.3)An unsupervised method based on multi-task Y-shaped network is proposed for nuclei segmentation of unlabeled pathological images.First,the source and target do-main datasets are processed and constructed.Then,domain adaptation training is ap-plied to the output space and feature space of the proposed multi-task Y-shaped net-work at the same time to fully learn the feature representation of the source domain.By comparing the results of other weakly supervised algorithms on the public dataset,the effectiveness of the proposed method is proved.Finally,based on the results of nu-clei segmentation,the nucleus morphology and distribution characteristics are extracted,and the lung pathological images are analyzed from the nucleus level,which proves the reliability of the method.