Deep Learning-based Neuronal Soma Segmentation For Light Microscope Images

Automated neuronal soma segmentation algorithm is an important part of neuron image analysis tools,it could extract important information from large-scale neuron image for research on brain function and brain disease instead of manual lableling manner efficicently,including neuron distribution and soma morphology.However,the limitation of existing automated neuronal soma segmentation algorithm makes them difficult to deal with some problems in soma segmentation for large-scale neuron image,including complex intensity variation,diverse soma morphology,touching somata,unrelated structure and so on.It is necessary to explore the design of automated neuronal soma segmentation algorithm,and build fast and accurate neuronal soma segmentation algorithm.This study explored the way to improve related aspects in accuracy and computational efficiency of neuronal soma segmentation based on deep learning,including:(1)Three automated soma segmentation methods were proposed for different applications,namely,Raybrust-UNet,Multi Task-UNet and Center-UNet.Raybrust-UNet was designed for Green Fluorescent Protein(GFP)labeled dataset,which combined 3D UNet for foreground extraction and a variant of Rayburst sampling algorithm for touching soma segmentation.The proposed method could predict accurate neuronal soma location and contour for GFP labeled dataset,achieving F1-score of 0.84 and average dice coefficient of 0.75.Mutli Task-UNet was designed for Nissl staining dataset,this method treat the soma segmentation task as contour dectecion and applied multi-task learning model to split touching soma and predict accurate soma contour.The results in GFP labeled dataset and Nissl staining demonstrated excellent performance in soma localization and segmentation.The F1-score and dice coefficient achieved 0.88 and 0.85,respectively.Finally,the CenterUNet was designed for fast neuronal soma segmentation in large-scale images,which adopted bounding box regression algorithm to segment neuronal soma in an End-to-End manner,which could segment soma segmentation accurately and quickly.This method achievied F1 score of 0.94 and dice coefficient of 0.79 in Nissl staining dataset,F1 score of0.90 and average dice coefficient of 0.77 in GFP labeled dataset.The proposed methods demonstrated better performance than other 5 existing neuronal soma segmentation methods.(2)a novel semi-supervised neuronal soma segmentation based on self-training was proposed to reduce manual labeling effort in large-scale datasets.The proposed method adopted self-training strategy to predict pseudo label for the unannotated training samples and improved the performance of soma segmentation model gradually.The initial training part and pseudo label updating method in self-training were modified to improved the performance of soma segmentation model.Results in different datasets demonstrated that the proposed method could be trained using training set with few manual annontations,the F1-score and dice coefficient achieved 0.93 and 0.77,respectively.And the performance is close to fully supervised soma segmentation model.(3)Neuronal soma segmentation method based on transfer learning was proposed for complex applications.This kind of method could build soma segmentation model using existing annontated training samples for a different dataset without annontations.The proposed method adopted unsupervised domain adaptation framework to improve the generalization of soma segmentation model.The results in different types of fluorescent protein labeled datasets demonstrated that the proposed method could reduce the impact of appearance difference between training set and testing set on soma segmentation performance.The proposed method demonstrated about 1% improvement in F1-score and could reduce the interference of neurite.Totally,this study proposed many kinds of automated neuronal soma segmentation methods,which demonstrated good performance in different types of datasets.Besides,this study proposed soma segemtation method based on semi-supervised and transfer learning to reduce manual labeling effort.After all,this study provides a group of effective deep learning tools for neuronal soma segmentation.