Research On Nuclear Segmentation Of Pathological Image Based On Convolutional Neural Network

Cancer has become the second leading cause of human death due to the increasing incidence and mortality of cancer worldwide.Cancer is a disease caused by molecular aberrations that cause normal cells to go out of control.It is characterized by the uncontrolled growth of cancer cells.Cancer cells not only grow rapidly at the primary site,but often metastasize and spread to other healthy parts of the body,furthering the disease.deterioration.In recent years,computer-aided diagnosis technology methods have been widely used in medical image analysis tasks.They use image analysis and computing technology to effectively process medical images,objectively evaluate obvious parts,and greatly improve the accuracy,speed and efficiency of medical diagnosis.level of automation.Although computer-aided diagnosis technology has played a huge role in medical diagnosis,the existing nuclear segmentation methods in histopathological images still have the problems of nuclear difference and nuclear adhesion,which limit the accuracy of nuclear segmentation.In cancer treatment,analyzing pathological cell nuclei and judging the state of cancer cells is crucial for cancer diagnosis.Therefore,overcoming these two problems and improving the accuracy of nucleus segmentation is worthy of further study by researchers.There are two problems in nuclei segmentation in histopathological images.On the one hand,the size,shape,and staining pattern of cancer nuclei vary greatly among different organs and cancer stages,and multiple differences make it difficult for pathologists to obtain consistent analysis results.On the other hand,the nuclei between the stained nuclei cluster regions have adhesions,and the boundaries between nuclei are blurred and difficult to distinguish,making the quantification of nuclei instances difficult.These factors pose great challenges for nucleus segmentation.In view of the above problems,this thesis takes the pathological image nucleus segmentation as the research goal,and carries out related research.This thesis treats the problem of nucleus segmentation as two related tasks: pixel-level segmentation(semantic segmentation)and instance-level segmentation(instance segmentation).Semantic segmentation classifies each pixel of the image,solves the problem of large differences in nuclear features,and enhances the objectivity of the segmentation results.Instance segmentation not only needs to identify the category of the pixel,but also needs to annotate the nucleus instance to which each nuclear pixel belongs,so each nucleus can be distinguished in the cohesive nucleus cluster and specific nucleus quantification results can be obtained.The main contributions and innovations of this thesis mainly include the following two parts:(1)According to the characteristics of nuclei in H&E-stained histopathological images,this study proposes a semantic segmentation model of nuclei based on edge points.In this thesis,the staining of pathological images is analyzed,and a strategy of propensity sampling for the edge of the nucleus is proposed according to the color feature difference between the nucleus and other parts.The nucleus is semantically segmented from the point of view of the edge,and the AJI index of 69.15% and the Dice index are obtained.The accuracy of 81.76% improves the effect of nuclear semantic segmentation.(2)This thesis proposes a cell nucleus instance segmentation model based on a deformable multi-level feature network,and studies the process of cell nucleus instance segmentation in the two stages of feature and segmentation.The code segmentation stage has been improved and enhanced,and the m AP and m AR on the Mo Nu Seg 2018 dataset reached 37.8% and 47.4%,respectively,which improved the effect of nucleus instance segmentation.