Segmentation of Non-Human Primate Cerebrovascular Images from Synchrotron Radiation Micro-Tomography Using Transfer Learning and Attention U-Net

doi:10.11871/jfdc.issn.2096-742X.2026.03.001

Abstract

Abstract:

[Objective] To address the challenges of scarce high-quality annotations and severe artifact interference in Synchrotron Radiation Micro-Tomography (SR-μCT) imaging of non-human primate cerebrovasculature, this study proposes an automated segmentation method based on transfer learning and a hierarchical combined weighting strategy. [Methods] A global percentile normalization strategy is employed to suppress artifacts while maintaining signal consistency. A “pre-training and fine-tuning” framework is established, innovatively incorporating a hierarchical combined weighting strategy to synergistically reinforce the capture of micro-vessels and boundary features. [Results] Experimental results demonstrate that the proposed method comprehensively outperforms nnU-Net 3D, achieving a Dice coefficient of 0.8686, a recall of 96.93%, and a topological connectivity metric (clDice) of 0.8848. These results signify a substantial resolution to the issues of micro-vessel miss-detection and disconnection. [Conclusions] This algorithm effectively overcomes the segmentation challenges associated with small-sample and high-resolution SR-μCT cerebrovascular imaging, laying a solid technical and data foundation for future large-scale sub-micron whole-brain imaging analysis and neurovascular network quantification in non-human primates and humans.

Key words: cerebrovascular segmentation, synchrotron radiation micro-tomography, transfer learning, hierarchical combined weighting, topological connectivity

YE Jing,WANG Chunpeng,LI Qintong,CHEN Zhuo,LI Zongze,ZHANG Jiaru,ZHANG Xiangzhi,HU Yuguang,TAI Renzhong. Segmentation of Non-Human Primate Cerebrovascular Images from Synchrotron Radiation Micro-Tomography Using Transfer Learning and Attention U-Net[J]. Frontiers of Data and Computing, 2026, 8(3): 1-14.

Figures/Tables 11

Fig.1

Fig.2

Fig.3

Fig.4

Fig.5

Fig.6

Table 1

Table 2

Table 3

Fig.7

Table 4

References 28

[1]	IADECOLA C. The Neurovascular Unit Coming of Age: A Journey through Neurovascular Coupling in Health and Disease[J]. Neuron, 2017, 96(1): 17-42. doi: S0896-6273(17)30652-9 pmid: 28957666
[2]	SWEENEY M D, SAGARE A P, ZLOKOVIC B V. Blood-brain barrier breakdown in Alzheimer disease and other neurodegenerative disorders[J]. Nature Reviews Neurology, 2018, 14(3): 133-150. doi: 10.1038/nrneurol.2017.188 pmid: 29377008
[3]	ZENG C, CHEN Z, YANG H, et al. Advanced high resolution three-dimensional imaging to visualize the cerebral neurovascular network in stroke[J]. International Journal of Biological Sciences, 2022, 18(2): 552-571. doi: 10.7150/ijbs.64373 pmid: 35002509
[4]	FILAN C, SONG H, PLATT M O, et al. Analysis of structural effects of sickle cell disease on brain vasculature of mice using three-dimensional quantitative phase imaging[J]. Journal of Biomedical Optics, 2023, 28(9): 096501.
[5]	LI H C, CHEN P Y, CHENG H F, et al. In Vivo Visualization of Brain Vasculature in Alzheimer’s Disease Mice by High-Frequency Micro-Doppler Imaging[J]. IEEE transactions on bio-medical engineering, 2019, 66(12): 3393-3401. doi: 10.1109/TBME.10
[6]	邰仁忠. X射线物理学[J]. 物理, 2021, 8(50): 501-511.
[7]	HWU Y, MARGARITONDO G. Synchrotron radiation and X-ray free-electron lasers (X-FELs) explained to all users, active and potential[J]. Journal of Synchrotron Radiation, 2021, 28(3): 1014-1029. doi: 10.1107/S1600577521003325
[8]	STAMPFL A P J, LIU Z, HU J, et al. SYNAPSE: An international roadmap to large brain imaging[J]. Physics Reports, 2023, 999: 1-60. doi: 10.1016/j.physrep.2022.11.003
[9]	CAO Y, ZHANG M, DING H, et al. Synchrotron radiation micro-tomography for high-resolution neurovascular network morphology investigation[J]. Journal of Synchrotron Radiation, 2019, 26(3): 607-618. doi: 10.1107/S1600577519003060
[10]	汤乔伟, 蔡小青, 李江, 等. 同步辐射X射线成像技术在脑成像研究中的应用[J]. 高等学校化学学报, 2022, 43(12): 128-145.
[11]	HESAMIAN M H, JIA W, HE X, et al. Deep Learning Techniques for Medical Image Segmentation: Achievements and Challenges[J]. Journal of Digital Imaging, 2019, 32(4): 582-596. doi: 10.1007/s10278-019-00227-x pmid: 31144149
[12]	RONNEBERGER O, FISCHER P, BROX T. U-Net: Convolutional Networks for Biomedical Image Segmentation[C]// Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015, 2015: 234- 241.
[13]	OKTAY O, SCHLEMPER J, FOLGOC L L, et al. Attention U-Net: Learning Where to Look for the Pancreas[J]. arXiv preprint arXiv:1804.03999, 2018.
[14]	FRANGI A F, NIESSEN W J, VINCKEN K L, et al. Multiscale vessel enhancement filtering[C]// Medical Image Computing and Computer-Assisted Intervention — MICCAI’98, 1998: 130-137.
[15]	CHEN X, WILLIAMS B M, VALLABHANENI S R, et al. Learning Active Contour Models for Medical Image Segmentation[C]// 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019: 11624-11632.
[16]	MANNIESING R, VELTHUIS B K, VAN LEEUWEN M S, et al. Level set based cerebral vasculature segmentation and diameter quantification in CT angiography[J]. Medical Image Analysis, 2006, 10(2): 200-214. pmid: 16263325
[17]	SALEH M M, SALIH M E, AHMED M A A, et al. From Traditional Methods to 3D U-Net: A Comprehensive Review of Brain Tumour Segmentation Techniques[J]. Journal of Biomedical Science and Engineering, 2025, 18(1): 1-32. doi: 10.4236/jbise.2025.181001
[18]	ÇIÇEK Ö, ABDULKADIR A, LIENKAMP S S, et al. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation[C]// Medical Image Computing and Computer-Assisted Intervention - MICCAI 2016, 2016: 424-432.
[19]	PARK M, OH S, PARK J, et al. ES-UNet: efficient 3D medical image segmentation with enhanced skip connections in 3D UNet[J]. BMC Medical Imaging, 2025, 25(1): 327. doi: 10.1186/s12880-025-01857-0 pmid: 40804359
[20]	ALWAN A H, ALI S A, HASHIM A T. Medical Image Segmentation Using Enhanced Residual U-Net Architecture[J]. Mathematical Modelling of Engineering Problems, 2024, 11(2): 507. doi: 10.18280/mmep
[21]	EHAB W, HUANG L, LI Y, et al. UNet and Variants for Medical Image Segmentation[J]. International Journal of Network Dynamics and Intelligence, 2024, 3(2): 100009.
[22]	GLÄNZER L, MASALKHI H E, ROETH A A, et al. Vessel Delineation Using U-Net: A Sparse Labeled Deep Learning Approach for Semantic Segmentation of Histological Images[J]. Cancers, 2023, 15(15): 3773. doi: 10.3390/cancers15153773
[23]	ALBERS J, SVETLOVE A, DUKE E. Synchrotron X-ray imaging of soft biological tissues-principles, applications and future prospects[J]. Journal of Cell Science, 2024, 137(20): jcs261953.
[24]	WITHERS P J, BOUMAN C, CARMIGNATO S, et al. X-ray computed tomography[J]. Nature Reviews Methods Primers, 2021, 1(1): 18. doi: 10.1038/s43586-021-00015-4
[25]	TRINKLE S, FOXLEY S, KASTHURI N, et al. Synchrotron x-ray microcomputed tomography as a validation dataset for diffusion MRI in whole mouse brain[J]. Magnetic resonance in medicine, 2021, 86(2): 1067-1076. doi: 10.1002/mrm.v86.2
[26]	HAMANN G F, DEL ZOPPO G J. Chapter 38-The Cerebral Microvasculature and Responses to Ischemia[M]// MOHRJ P, CHOID W, GROTTAJ C, et al. Stroke (Fourth Edition). Philadelphia: Churchill Livingstone, 2004: 775-783.
[27]	TETTEH G, EFREMOV V, FORKERT N D, et al. DeepVesselNet: Vessel Segmentation, Centerline Prediction, and Bifurcation Detection in 3-D Angiographic Volumes[J]. Frontiers in Neuroscience, 2020, 14: 592352. doi: 10.3389/fnins.2020.592352
[28]	ISENSEE F, JAEGER P F, KOHL S A A, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation[J]. Nature methods, 2021, 18(2): 203-211. doi: 10.1038/s41592-020-01008-z pmid: 33288961

参数类别	参数名称	预训练阶段	微调阶段
数据参数	输入图像尺寸（像素）（padding之前）	325×304	325×304
	训练/验证集划分	7,050/1,050	5,120/1,280（5折交叉验证）
	测试集数量（张）	16,800	1,600
训练超参数	批量大小（Batch Size）	24	64
	优化器（Optimizer）	Adam	Adam
	训练周期（Epochs）	50	20
	初始学习率（Learning Rate）	1×10^-5	1×10^-4
	学习率衰减	Val Loss plateau (patience=5)	Val Loss plateau (patience=5)
推理参数	分割预测阈值（Threshold）	0.5	0.5

模型	Dice(↑)	IoU(↑)	HD95(pixel) (↓)	clDice(↑)	Recall(↑)	Precision (↑)	F2 Score (↑)
nnU-Net 2D	0.7871±0.0773	0.6543±0.1048	1.6293±0.4265	0.7583±0.0372	0.7392±0.1291	0.8633±0.0953	0.7559±0.1068
nnU-Net 3D	0.8403±0.0686	0.7293±0.0999	1.3949±0.2602	0.7801±0.0375	0.8245±0.0983	0.8614±0.0586	0.8303±0.0855
本文方法	0.8686±0.0191	0.7681±0.0299	1.2485±0.2269	0.8848±0.0194	0.9693±0.0131	0.7869±0.0236	0.9263±0.0155

训练策略	预训练	小目标权重	边界权重	Dice (↑)	IoU (↑)	HD95 (pixel) (↓)	clDice (↑)	Recall (↑)	Precision (↑)	F2 Score (↑)
1	×	√	√	0.8516	0.7419	1.4142	0.8752	0.9645	0.7624	0.9159
2	√	×	×	0.8589	0.7531	1.3314	0.8004	0.9399	0.7907	0.9057
3	√	×	√	0.8407	0.7262	2.4152	0.7213	0.9064	0.7841	0.8789
4	√	√	×	0.8032	0.6722	1.5413	0.8510	0.9656	0.6881	0.8932
本文	√	√	√	0.8709	0.7717	1.2485	0.8702	0.9637	0.7944	0.9243

模型	参数量（M）	计算复杂度（G-MACs）	推理模式	耗时（ms/步）	吞吐量（FPS）
nnU-Net 2D	33.47	71.90	单张	4.36	229
nnU-Net 3D	31.20	1,228.40	滑动窗口	31.70^*	31
本文方法	34.88	103.85	单张	5.48	182