摘要
血管分割是生物医学图像处理中常见的实验操作,是准确分析血管成像数据的基础。针对现有基于监督学习血管分割方法面临的标签数据难以获取、普适性差的问题,提出一种基于域迁移的弱监督学习血管图像分割实验方法。该方法基于循环生成对该抗网络进行无监督血管风格迁移,并利用极少量标签数据对网络进行监督优化。利用三种不同显微光学血管图像对该方法进行性能评估,结果表明,在仅使用少量标签数据的情况下,分割结果F1值超过85%,血管长度密度差异和分叉点密度差异均低于10%,且能够用于大范围血管分析任务。该方法为血管分割提供了一种新的解决方案,提高了实验效率。
[Objective]Microscopic optical imaging enables the acquisition of capillary-resolution vascular network images,providing essential experimental data support for vascular research.Vessel segmentation,as a critical experimental step in biomedical image processing,is fundamental for accurately analyzing vascular imaging data.However,manual vessel segmentation methods are time-consuming and inaccuracy prone,while existing supervised learning-based vessel segmentation methods face challenges such as the difficulty of obtaining labeled data and limited generalizability.To address these issues,this study proposes a vessel segmentation method based on weakly supervised learning with domain transfer.[Methods]This method first employs a cycle-consistent generative adversarial network to achieve unsupervised style transfer from grayscale vascular images in the source domain to binary vascular images in the target domain without requiring paired images and annotated data.On this basis,a single binary-labeled vascular image block is used for supervised learning to constrain the generator’s training process.The generator uses 3D U-Net as the backbone network,while the discriminator adopts a 3D fully convolutional neural network structure.The network training follows an alternating strategy:In odd-numbered iterations,the generator and discriminator are trained using cycle consistency loss and adversarial loss.The mean squared error loss function and smooth L1 loss function are used for adversarial and cycle consistency loss,respectively.In even-numbered iterations,the single annotated data block optimizes the generator model parameters.The Dice loss function is employed as the supervised learning loss to correct errors between the segmented image and the ground truth label,allowing the generator to better learn the data distribution and features.Furthermore,in response to researchers’interest in microvascular density and other statistical parameters,two new evaluation metrics are introduced to quantitatively assess segmentation performance:the branching point density difference percentage and the length density difference percentage.These metrics effectively measure the impact of vessel segmentation results on vascular statistical parameters.[Results]Performance evaluation of the method was conducted using mouse brain vascular images acquired from three microscopic optical imaging techniques:high-definition fluorescent micro-optical sectioning tomography(HD-fMOST),light sheet microscopy,and fluorescent micro-optical sectioning tomography.The results demonstrate that the proposed method performs considerably with only minimal labeled data.The F1-score of the segmentation results exceeds 85%,and the differences in vascular length density and bifurcation point density are less than 10%.Compared to existing methods,the proposed method achieves more accurate segmentation of vascular topological structures.Moreover,ablation experiments validate the effectiveness of the unsupervised style transfer module and the single data block supervised learning module in the proposed method.Finally,large-scale segmentation of vascular images from the motor cortex of mice imaged using HD-fMOST was conducted,followed by reconstruction and analysis,further confirming the method’s effectiveness and reliability in practical applications.[Conclusions]The proposed weakly supervised learning-based vessel segmentation method achieves accurate segmentation of microscopic optical vascular images with minimal labeled data.It not only considerably reduces data annotation costs and improves experimental efficiency but also provides a new experimental approach for vessel segmentation and analysis,offering substantial practical value in vascular research.
作者
李宇昕
王鑫龙
李军怀
张仟龙
李安安
LI Yuxin;WANG Xinlong;LI Junhuai;ZHANG Qianlong;LI Anan(School of Computer Science and Engineering,Xi’an University of Technology,Xi’an 710048,China;Wuhan National Laboratory for Optoelectronics,Huazhong University of Science and Technology,Wuhan 430074,China)
出处
《实验技术与管理》
北大核心
2025年第5期45-53,共9页
Experimental Technology and Management
基金
科技部科技创新2030重大项目(2021ZD0201002)
国家自然科学基金项目(82102137)
华中科技大学武汉光电国家研究中心开放基金(2021WNLOKF006)。
关键词
血管图像分割
显微光学成像
域迁移
弱监督学习
生物医学图像
vascular image segmentaion
microscopic optical imaging
domain transfer
weakly supervised learning
biomedical image
