Research On Tissue Segmentation Of Digital Pathology Image Of Non-small Cell Lung Cancer

Lung cancer is one of the most common malignant tumors in the world.It ranks first in morbidity and mortality of all the malignant tumors,which pose a serious threat to the life and health of the nation.Non-small cell lung cancer(NSCLC)is the most common histological type of lung cancer,accounting for about 85% of all lung cancer cases.Pathological diagnosis is the gold standard of clinical medicine in which especially the assessment of tumor cells,tumorinfiltrating lymphocytes,stroma and necrosis,can serve as the basis for further diagnosis and analysis of NSCLC.With the rapid development of digital pathology and artificial intelligence technology,intelligent analysis of Whole Slide Images(WSI)has become the tendency of pathological diagnosis.Accurate segmentation of various tissues is a key step to realize automatic,accurate and intelligent analysis of WSIs.After making remarkable achievements in the field of image segmentation,Deep Learning(DL)and Convolutional Neural Network(CNN)are the most effective methods for segmentation tasks in recent years.Currently,DCNN-based fully supervised image segmentation methods rely on dense pixel-level labels.In order to realize tissue segmentation automatically and reduce the dependence of data labeling,we designed a weakly supervised semantic segmentation method for segmentation of four kinds of tissues in WSI,including tumor,lymphocyte,stroma and necrosis.In this study,we carry out in-depth research on weakly supervised tissue segmentation of pathological image from two aspects:(1)Weakly-supervised Tissue Semantic Segmentation with Curriculum Self-supervised Learning.In this research,we train the CNN with image-level labels and acquire the segmentation results based on Class Activation Maps(CMAs).In order to make up for the weak supervision information,we propose a Curriculum Self-supervised Learning(CSSL)strategy,which force the CNN to discover the internal characteristics of the original data without any additional labels.We designed three subtasks of CSSL,including reconstruction,inpainting and color deconvolution.Then,the difficulty of self-supervised learning tasks is increased by superimposing subtasks,so that CNN can learn a more abundant feature representation,which will have a more positive effect on the main segmentation task.(2)Distinctness Dropout and Multi-supervision Based Method for Weakly-supervised Tissue Semantic Segmentation.It is easy the CAM-based weakly supervised methods identify only the most discriminative part of the target object,without the capacity of covering entire extent of the object.In response to this problem,we proposed a CAMD(Class Activation Maps based Dropout)module,which randomly discards high-response regions in the feature maps to force DCNN to explore the less discriminative parts of the target object.In addition,drawing on the idea of noisy labels and multi supervision,we utilize the feature maps of different convolutional layers of DCNN to generate multiple sets of pseudo labels,and then use them to train a fully supervised semantic segmentation network,which reduces the influence of noisy labels on network training.In this study,we also developed a dataset for weakly supervised tissue segmentation of NSCLC WSIs.The training set includes 16678 image-level annotated pathology images,and the test set contains 607 pixel-level annotated pathology images.Based on this dataset,we conducted detailed experiments validation and analysis on the methods proposed above.Experiment results show that the proposed method can effectively improve the performance of tissue segmentation of NSCLC pathological images.