1.北京工业大学化学与生命科学学院(北京 100124)
2.强联智创(北京)科技有限公司(北京 100176)
高斌。E-mail:gaobin@bjut.edu.cn
收稿:2025-04-24,
修回:2025-08-11,
网络首发:2026-08-17,
纸质出版:2026-08-20
移动端阅览
李中凯,李晓琴,高斌,等.一种基于残差多尺度注意力生成对抗网络的颅内动脉瘤血流动力学预测方法[J].北京生物医学工程,2026,45(4):363-370.
LI Zhongkai,LI Xiaoqin,GAO Bin,et al.A hemodynamic prediction method for intracranial aneurysms based on residual multi-scale attention generative adversarial network[J].Beijing Biomedical Engineering,2026,45(4):363-370.
李中凯,李晓琴,高斌,等.一种基于残差多尺度注意力生成对抗网络的颅内动脉瘤血流动力学预测方法[J].北京生物医学工程,2026,45(4):363-370. DOI: 10.3969/j.issn.1002-3208.2026.04.005.
LI Zhongkai,LI Xiaoqin,GAO Bin,et al.A hemodynamic prediction method for intracranial aneurysms based on residual multi-scale attention generative adversarial network[J].Beijing Biomedical Engineering,2026,45(4):363-370. DOI: 10.3969/j.issn.1002-3208.2026.04.005.
目的
2
颅内动脉瘤是一种严重威胁患者生命安全的疾病,其速度、压力和涡量的血流动力学预测对于诊断和治疗具有重要的临床意义。计算流体力学(computational fluid dynamics, CFD)能够有效地模拟血流动力学特征,但其计算成本较高。为此本文建立一种新型的深度学习架构,用于快速预测颅内动脉瘤的血流动力学。
方法
2
基于210例颅内动脉瘤患者的形状数据进行计算流体力学模拟,并根据计算结果构建压力、速度和涡量数据库。设计了一种结合了残差-挤压激励-模块(Res-SE-Block)和多尺度注意块(multi-scale attention block,MAB)的残差多尺度注意力生成对抗网络(residual multi-scale attention generative adversarial network,RMAGAN),基于动脉瘤的形状信息来进行压力、速度和涡量的血流动力学预测。生成器采用了残差多尺度U-Net架构,用以提高网络的性能和预测能力,判别器采用了马尔可夫判别器,来提高网络的运行速度,损失函数结合了CGAN的损失函数与Smooth损失函数。最后使用均方根误差(root mean square error, RMSE)、归一化平均绝对误差(normalized mean absolute error, NMAE)和皮尔逊相关系数(
r
)指标对本文的深度学习模型进行评估,同时与多层感知机和卷积神经网络进行了比较。
结果
2
深度学习模型的预测结果与计算流体力学模拟的速度、压力和涡量非常吻合,两种方法的相似度和准确性达到RMSE
<
0.280 4, NMAE
<
12.509 6%,
r
>
0.815 3,同时减少了1 500倍的计算成本。
结论
2
本文的深度学习模型能够进行血流动力学的快速预测,并且大幅降低计算成本,为血流动力学模拟提供了一种全新的解决思路,在颅内动脉瘤的临床诊断和治疗研究中具有潜力。
Objective
2
Intracranial aneurysm is a disease that seriously threatens the life and safety of patients, and the hemodynamic prediction of velocity, pressure, and vorticity holds clinical significance for their diagnosis and treatment. Computational fluid dynamics (CFD) effectively models hemodynamic characteristics, it incurs high computational costs. The study proposes a novel deep learning architecture for the rapid prediction of intracranial aneurysm hemodynamics.
Methods
2
The CFD simulations were conducted based on the shape data of 210 intracranial aneurysm cases, and a database of pressure, velocity, and vorticity was established according to the simulation results. A residual multi-scale attention generative adversarial network (RMAGAN) combining residual squeeze-and-excitation blocks (Res-SE-Block) and multi-scale attention block (MAB) was designed to perform hemodynamic prediction of pressure, velocity and vorticality based on the shape information of aneurysms. The generator employed a residual multi-scale U-Net architecture to enhance network performance and predictive capability, while the discriminator utilized a Markov discriminator to improve computational efficiency. The loss function integrates the CGAN loss function with a Smooth loss function to optimize training stability and prediction accuracy. The performance of our deep learning model was evaluated by using root mean square error (RMSE), normalized mean absolute error (NMAE), and Pearson correlation coefficient(
r
), and was compared with that of multi-layer perceptron (MLP) and convolutional neural network(CNN).
Results
2
The predicted results of the deep learning model closely matched the velocity, pressure, and vorticity from the CFD simulations. The similarity and accuracy between the two methods achieved RMSE
<
0.280 4, NMAE
<
12.509 6%, and
r
>
0.815 3, while reducing the computational cost by a factor of 1 500.
Conclusions
2
Our deep learning model enables rapid prediction of hemodynamics and significantly reduces computational costs, providing a novel solution for hemodynamic simulations. It holds potential for clinical diagnosis and treatment of intracranial aneurysms.
Xu Z, Rui YN, Hagan JP, et al. Intracranial aneurysms: pathology, genetics, and molecular mechanisms[J]. Neuromolecular Medicine, 2019, 21(4): 325-343.
Liang F, Oshima M, Huang H, et al. Numerical study of cerebroarterial hemodynamic changes following carotid artery operation: a comparison between multiscale modeling and stand-alone three-dimensional modeling[J]. Journal of Biomechanical Engineering, 2015, 137(10): 101011.
Fu W, Gu Z, Meng X, et al. Numerical simulation of hemodynamics in stented internal carotid aneurysm based on patient-specific model[J]. Journal of Biomechanics, 2010, 43(7): 1337-1342.
Liang L, Liu M, Martin C, et al. A deep learning approach to estimate stress distribution: a fast and accurate surrogate of finite-element analysis[J]. Journal of the Royal Society, Interface, 2018, 15(138): 20170844.
Noorbakhsh-Sabet N, Zand R, Zhang Y, et al. Artificial intelligence transforms the future of health care[J]. The American Journal of Medicine, 2019, 132(7): 795-801.
Li G, Wang H, Zhang M, et al. Prediction of 3D cardiovascular hemodynamics before and after coronary artery bypass surgery via deep learning[J]. Communications Biology, 2021, 4(1): 99.
Itu L, Rapaka S, Passerini T, et al. A machine-learning approach for computation of fractional flow reserve from coronary computed tomography[J]. Journal of Applied Physiology (1985), 2016, 121(1): 42-52.
Gharleghi R, Sowmya A, Beier S. Transient wall shear stress estimation in coronary bifurcations using convolutional neural networks[J]. Computer Methods and Programs in Biomedicine, 2022, 225: 107013.
Su B, Zhang JM, Zou H, et al. Generating wall shear stress for coronary artery in real-time using neural networks: feasibility and initial results based on idealized models[J]. Computers in Biology and Medicine, 2020, 126: 104038.
Lv L, Li H, Wu Z, et al. An artificial intelligence-based platform for automatically estimating time-averaged wall shear stress in the ascending aorta[J]. European Heart Journal- Digital Health, 2022, 3(4): 525-534.
Wang S, Wu D, Li G, et al. Deep learning-based hemodynamic prediction of carotid artery stenosis before and after surgical treatments[J]. Frontiers in Physiology, 2023, 13: 1094743.
Morales Ferez X, Mill J, Juhl KA, et al. Deep learning framework for real-time estimation of in-silico thrombotic risk indices in the left atrial appendage[J]. Frontiers in Physiology, 2021, 12: 694945.
Liang L, Mao W, Sun W. A feasibility study of deep learning for predicting hemodynamics of human thoracic aorta[J]. Journal of Biomechanics, 2020, 99: 109544.
Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial nets[C]// 28th International Conference on Neural Information Processing Systems. Montreal: NIPS, 2014, 2: 2672-2680.
P J, B K SK, Jayaraman S. Automatic foot ulcer segmentation using conditional generative adversarial network (AFSegGAN): a wound management system[J]. PLOS Digital Health, 2023, 2(11): e0000344.
Isola P, Zhu J, Zhou T, et al. Image-to-image translation with conditional adversarial networks[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI, USA: CVPR, 2017: 5967-5976.
van der Schot A, Sikkel E, Niekolaas M, et al. Placental vessel segmentation using pix2pix compared to U-Net[J]. Journal of Imaging, 2023, 9(10): 226.
Gharahi H, Zambrano BA, Zhu DC, et al. Computational fluid dynamic simulation of human carotid artery bifurcation based on anatomy and volumetric blood flow rate measured with magnetic resonance imaging[J]. International Journal of Advances in Engineering Sciences and Applied Mathematics, 2016, 8(1): 40-60.
Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation[C]// MICCAI 2015. Munich, Germany: Cham: Springer International Publishing, 2015: 234-241.
Jiang L, Ou J, Liu R, et al. RMAU-Net: residual multi-scale attention U-Net for liver and tumor segmentation in CT images[J]. Computers in Biology and Medicine,2023,158: 106838.
Girshick BR. Fast R-CNN[C]// 2015 IEEE International Conference on Computer Vision (ICCV). Santiago, Chile: ICCV, 2015: 1440-1448.
0
浏览量
0
下载量
0
CSCD
关联资源
相关文章
相关作者
相关机构
京公网安备11010802024621