A Semi-supervised Nuclear Instance Segmentation Method Based On Anchor Offset Prediction

Automatic analysis of medical microscopic images to determine the location of nuclei is an important stage in tumor diagnosis and cancer detection.Nucleus segmentation and quantification will help disease researchers analyze the causes,pathogenesis and development process of lesions,so as to make pathological diagnosis.The traditional nucleus instance segmentation algorithm requires label complete data,but there are a large number of nuclei in medical microscopic images,the high heterogeneity of nuclei in different tissues,and the nucleus boundary tends to be blurred in the case of defocus.For many reasons,the labeling cost is very high,and it is difficult to label all instances of nucleus instances.In order to solve the above contradiction,this paper proposes a semi-supervised nuclear instance segmentation method based on anchor offset prediction(AOP-Mask),which realizes nucleus instance segmentation when only partial nucleus instances are labeled,and makes a contribution to alleviating the problems of high cost and time-consuming analysis of medical data labeling.The experimental results show that AOP-Mask is better than Mask RCNN in many evaluation indexes,and reaches SOTA on partially labeled Mo Nu Seg dataset.The improvement scheme is as follows:(1)CBAM is applied to FPN.In order to make the model focus on the currently detected nucleus itself rather than other nuclei around it.We add CBAM to the feature output position of the FPN,so that the model can learn to forwardly focus on these key information.(2)Anchor offset prediction branch(AOP branch)is proposed.In view of the problem that the input features of the second stage regression branch lose part of the nucleus feature information during the down sampling of Roi Align,and the detection frame with better precision may be suppressed by the detection result with lower accuracy in the post-processing of the non maximum suppression,we propose the anchor offset prediction branch to replace the regression branch in the second stage,while avoiding the loss of nucleus feature information caused by down sampling,the predicted anchor offset is used to assist the non maximum suppression process,so as to retain the detection frame with better detection result.(3)A semi-supervised training method based on pseudo label merging is proposed.In view of the high cost of complete labeling of nuclei in medical images and the decline in the accuracy of the traditional nucleus instance segmentation model when using only partially labeled dataset for training,we propose a two-stage semi-supervised training method based on pseudo label generation and merging.This method can alleviate the imbalance between positive and negative samples caused by only partial nucleus labeled in the dataset,and achieve the result of higher precision nucleus instance segmentation in the case of partial labeling.