基于深度学习的轴突三维图像分割与重构  

Deep-learning based three-dimensional axon image segmentation and reconstruction

作　　者：黄天怡 冉小炜 郭增才[1,2,3] Huang Tianyi;Ran Xiaowei;Guo Zengcai(Tsinghua-Peking Joint Center for Life Sciences,Tsinghua University,Beijing 100084,China;IDG/McGovern Institute for Brain Research,Tsinghua University,Beijing 100084,China;Department of Basic Medical Sciences,School of Medicine,Tsinghua University,Beijing 100084,China;Department of Biomedical Engineering,School of Medicine,Tsinghua University,Beijing 100084,China)

机构地区：[1]清华大学生命科学联合中心,北京100084 [2]清华大学清华-IDG/麦戈文脑科学研究院,北京100084 [3]清华大学医学院基础医学系,北京100084 [4]清华大学医学院生物医学工程系,北京100084

出　　处：《神经解剖学杂志》2020年第3期261-269,共9页Chinese Journal of Neuroanatomy

基　　金：清华大学IDG/麦戈文“Brain+X”种子基金。

摘　　要：目的:设计适用于神经轴突荧光显微图像三维分割任务的深度神经网络(DNN),提高神经轴突自动重构的准确性.方法:将三维图像分割任务转化为在3个正交投影方向上的二维分割任务,在开源网络Deep-MACT的基础上,利用多通道输入降低压缩维度上的信息损失,利用神经轴突骨架加权的损失函数训练以强调神经骨架连贯性,提出SWCUnet.采用转盘共聚焦荧光显微成像采集的小鼠大脑稀疏标记神经元神经轴突部分49个图块作为数据集,以人工标注的形态重构结果作为金标准训练网络.对网络模型输出的二维分割图像进行三维复原,并以三维分割结果输入MOST算法进行自动重构.结果:SWCUnet(32通道,骨架权重5)三维分割F1-score达到0.662,较DeepMACT提升0.132.基于三维分割结果的自动重构F1-score达到0.80,比基于原图的自动重构提升0.24.结论:SWCUnet可以较好地提取高分辨荧光显微图像中的神经轴突特征,输出的三维分割结果实现了大幅提升神经轴突骨架形态的自动重构准确率的目标,为小鼠大脑稀疏标记神经元全脑成像数据的大规模自动化重构提供了一种新工具.Objective: Design a deep neural network(DNN) suitable for axon segmentation tasks based on 3D fluorescence microscopic images,to improve the accuracy of segmentation-based axon auto-reconstruction.Methods: 3D segmentation tasks are first transformed into 2D segmentations of their projections along the three orthogonal axis as our baseline model,the deep learning-enabled metastasis analysis in cleared tissue(Deep MACT),requires.On the basis of that,we propose the skeleton weighted multi-channel Unet(SWCUnet) with a newly designed multi-channel input strategy reducing spatial information loss,and a skeleton-weighted loss function facilitating continuity.Data-set includes 49 sparse-axon-containing image volumes of the sparse-labeling mouse brain imaged by spinning-disc confocal fluorescence microscopy,and manually annotated axon reconstructions serve as ground truth labels in network training.Finally,3D segmentations are recovered from their three orthogonal projections’ prediction generated by our network,and then fed into the micro-optical sectioning tomography(MOST) algorithm for auto-reconstruction.Results: SWCUnet(32-channel,skeleton weight 5) achieves 0.662 in F1-score for 3D segmentations,which is 0.132 higher than Deep MACT.Auto-reconstruction based on these 3D segmentations reaches 0.80 in node-based F1-score,which is 0.24 higher than raw image based auto-reconstruction.Conclusion: SWCUnet performs well in extracting axonal features from high resolution fluorescence microscopic images,and produces 3D segmentations that fulfill the aim of improving the accuracy of axon skeleton structure auto-reconstruction with great significance.Therefore,SWCUnet makes a new tool for large-scale automatic reconstruction of sparsely labeled neurons in whole-mouse-brain imaging data-sets.

关 键 词：深度学习 神经轴突 图像分割 重构 SWCUnet 小鼠 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]