| Pathology is the cornerstone of modern medicine,especially cancer treatment,and with the advent of Whole-Slide Images(WSI),digital pathology has also emerged.At present,the automatic analysis and computer-aided diagnosis of digital pathology images mainly suffer from problems such as scarcity of annotation data,complex prediction requirements,and diversified targets.This paper first proposes a semi-automatic medical image annotation method,which provides sufficient annotation data for subsequent model training.To automatically analyze PD-L1 stained slides,it proposes a multi-stage fully-supervised framework for pathology image interpretation with multiple prediction functions such as segmentation,classification,counting,and scoring of lung squamous cell carcinoma cells.Finally,this paper presents a learning algorithm which can exploit sparse point labels and realize weakly supervised segmentation.The main research content and work of this paper are as follows:First,to handle the problem that the labeling of pathological images is highly professional and the workload is scarce,this paper proposes a new semi-automatic medical image labeling method and a web application,which provides annotations for training machine learning models.Compared with existing labeling systems,this application is equipped with free drawing and labeling,communication and collaboration mechanism,and neural network auxiliary labeling,which can better adapt to the particularity and challenges of medical image labeling,and effectively improve the accuracy and efficiency of labeling collection,which alleviates the problem of scarcity of manual annotations.Second,to handle the problem that pathological image prediction needs are complex and a single model is difficult to effectively deal with,this paper proposes a multi-stage fully supervised pathology image interpretation framework based on deep learning.The framework includes three stages,namely tumor area segmentation,tumor cell segmentation,and positive and negative tumor cell classification.Finally,in the segmentation and classification results,the number of tumor cells and the Tumor Proportion Score(TPS)are calculated by the connected region labeling method.The pixel-level classification accuracy of this framework on the pathological slides of lung squamous cell carcinoma stained by PD-L1 immunohistochemical staining can reach more than 90%,and the TPS estimation error of most samples is within 5%.It provides a solution to the complex prediction requirements in pathological image processing.Third,due to differences in tissue morphology,staining methods and scanning equipment,etc.,pathology images exhibit obvious target morphological diversity.Therefore,this paper proposes a weakly supervised histopathology image segmentation algorithm,which uses only sparse point labels to train a semantic segmentation model with good generalization.Through hierarchical feature expression based on deep neural networks,the algorithm can identify pathological tissues with large differences from complex backgrounds,and embed dynamic label propagation into the end-to-end training framework,learn the inherent manifold structure of pathological images,and improve generalizability.This algorithm alleviates the problem of lack of manual marking from the perspective of computational models,and at the same time can effectively deal with the problem of target diversity in pathological images,indicating good application value.In summary,the annotation method and the algorithm models proposed in this research are expected to solve the problems of scarcity of manual annotations,complex prediction requirements,and diversification of pathological targets in digital pathology image analysis.They can enhance the efficiency and accuracy in automatic interpretation of pathology images,upgrade the level of computer-aided diagnosis,and promotes researches and developments in digital pathology,so as to provide patients with better medical services. |
PDF Full Text Request