Research On Contour-Based Nuclei Instance Segmentation

Data published in 2020 shown that cancer is the second leading cause of death in the world,and about one-sixth of all deaths in the world are caused by cancer.The clinical diagnosis of many cancers is mainly through the organization of pathological assessments in China.The diagnosis of cancer cases has increased every year,and the support life cycle of pathologists has led to the diagnosis of pathological diseases.With the rapid development of deep learning,it is possible to diagnose diseases with the help of intelligent calculation.This is an important step,because the classification and analysis of cancer mainly depends on the quality of nuclei segmentation.In this article,combining the characteristics of pathological images and the segmentation algorithm of cell nucleus instance based on studying contours,the existing algorithm has been improved.The main framework used in this article is a single-stage instance segmentation algorithm based on Polar Mask.Polar Mask transforms the problem of instance segmentation into predicting the instance center and the distance between the instance center and the contour in polar coordinates.Although it is a simple and effective instance segmentation algorithm,this method has two disadvantages when predicting each contour point: feature misalignment and insufficient spatial context information.For this reason,this article focuses on solving the above two problems.Aiming at the problem of feature misalignment,this paper uses a coarse-to-fine paradigm to redesign the contour regression branch,and introduces offset learning to further fine-tune the rough segmentation results.In order to reduce the amount of parameters and increase the speed of inference,the form of ray grouping is adopted to further compress the network structure under the premise of ensuring performance.In addition,a hierarchical grouping sampling method is used to distinguish between high-semantic and high-resolution features to improve the ability to reuse features.In dealing with the problem of insufficient spatial context information,this article introduces shape prior knowledge through graph convolution,circular convolution and coordinate convolution based on the circular structure of the cell nucleus,and uses the context information of the local area to fine-tune the rays.Finally,the cell topology map is constructed on the basis of instance segmentation,and the interpretability analysis of characteristic cancer types is performed to assist doctors in pathological diagnosis.This article mainly analyzes the importance of nodes in the cell topology graph from the perspectives of layerwise relevance propagation(LRP)and class activation map(CAM).