Rapid and automatic hemodynamic assessment: integration of deep learning-based image segmentation, vessel reconstruction, and CFD prediction

Introduction

Cardiovascular diseases affect millions of individuals worldwide, and thus represent a significant public health concern (1,2). Prompt diagnosis and proper treatment can significantly reduce symptoms, and improve patients’ chances of recovery. Hemodynamic analysis is a scientific discipline that studies blood flow in the circulatory system. Hemodynamic simulations provide critical hemodynamic parameters that cannot be detected using conventional clinical tests, such as wall shear stress (WSS), and the shear oscillation coefficient (3). Integrating hemodynamic analysis into the diagnostic process has the potential to enhance the accuracy of cardiovascular disease diagnosis. Thus, hemodynamic analysis has attracted considerable attention from physicians and researchers as an auxiliary diagnostic tool.

In current practice, vascular hemodynamic analyses are commonly performed using commercial software. This process generally involves two primary steps. First, the three-dimensional (3D) geometry of the vessel is reconstructed from medical images. Second, a computational mesh is generated, and a computational fluid dynamics (CFD) analysis is then conducted. Zhang et al. (4) conducted a study using Mimics to reconstruct three aortic models, focusing on the relationship between the type B aortic dissection mechanism in the longitudinal expansion direction and the hemodynamic parameters. Similarly, Condemi et al. (5) used 3D Slicer to segment the region of interest (ROI), and conducted numerical simulations using COMSOL to study popliteal artery entrapment syndrome. Hyde-Linaker et al. (6) employed the open-source software ITK-SNAP for 3D reconstruction to investigate hemodynamic characteristic changes of arteriovenous fistula before and after surgery. Soliveri et al. (7) used the Vascular Modeling ToolKit for arteriovenous fistula reconstruction, and examined the influence of different morphologies on hemodynamics. Abhilash et al. (8) conducted a 3D reconstruction of a bifurcated carotid artery model using Mimics to investigate the effect of bifurcation on hemodynamics. Kang et al. (9) leveraged the open-source software package SimVascular to model and analyze the influence of carotid atherosclerosis on cerebral hemodynamics. Caddy et al. (10) employed CFD simulations to examine hemodynamic changes in a 3D arterial network connecting the heart and eyes under microgravity conditions. Their methodology involved the manual segmentation of small arteries using the open-source image processing software GIMP, followed by further segmentation in Mimics to create a detailed 3D geometric model. Luan et al. (11) used Mimics for the 3D reconstruction of the aorta, employing CFD and morphological methods to explore the causes of distal stent graft-induced new entry tear during thoracic endovascular aortic repair for type B aortic dissection. Their findings suggest that specific morphological features and hemodynamic changes in the distal stented aorta may significantly contribute to distal stent graft-induced new entry tear.

It is essential to note that reconstructing 3D cardiovascular models using medical image processing software requires proficient medical personnel to manually annotate the ROIs. Further, patient examinations may yield several hundred computed tomography (CT) slices, and the high gray-scale values of CT images can create challenges for medical professionals, increasing the likelihood of errors in judgment in demanding work environments. Such errors can significantly affect the outcomes of 3D modeling. Moreover, medical professionals must have a strong CFD foundation to conduct hemodynamic simulations on reconstructed vascular models.

Recently, numerous advanced segmentation technologies have emerged due to the evolution of medical image segmentation algorithms. These algorithms can automatically identify ROIs in medical images, thereby assisting medical professionals by minimizing the repetitive and tedious tasks in the initial stages. For example, Colombo et al. (12) used an enhanced region-growing algorithm to segment and reconstruct lower-limb arteries. Cao et al. (13) employed a convolutional neural network (CNN) to segment aortic dissections. Duan et al. (14) used a UNet for the preliminary localization and segmentation of coronary arteries, followed by precise segmentation and extraction using the level set algorithm. Chen et al. (15) introduced a 3D medical image reconstruction platform that leverages a sequence of two-dimensional (2D) CT images to generate 3D models. This approach served as a fundamental step in the progression of medical image 3D reconstruction, as it both enhances surgical precision and minimizes tissue trauma. Yin et al. (16) pioneered the remote 3D reconstruction of medical images in a browser/server framework. Using VTK and HTML5, they achieved the remote display and interaction of medical image 3D reconstruction. They also proposed a method to display 3D images on a pure web client, and exhibited its real-time performance on both PC and mobile platforms. Further, Mittal et al. (17) applied the modified-Balanced Iterative Reducing and Clustering using Hierarchies algorithm to brain tumor segmentation, subsequently employing the marching cubes (MC) algorithm for 3D reconstruction. These artificial intelligence-based algorithms have the potential to achieve the automatic segmentation and 3D reconstruction of blood vessels, leading to a substantial reduction in workload.

Nonetheless, research that combines CFD with deep learning-based automatic vascular modeling is currently limited. However, the amalgamation of these disciplines holds promise in automating the analysis of blood flow patterns, providing rapid and user-friendly outcomes for medical practitioners. This would enable fully automatic hemodynamic analysis without human intervention, and provide expeditious and intuitive results for medical professionals.

To address the issue of automating blood flow analysis, the present study introduced a novel method for analyzing vascular hemodynamics. This approach uses a deep learning-based image segmentation algorithm alongside CFD. Specifically, CNNs are employed to automatically segment 2D CT image slices, and the resulting segments are used in tandem with the MC algorithm for 3D reconstruction. OpenFOAM, a versatile and customizable open-source CFD software known for its cross-platform support, was selected to conduct hemodynamic simulations on the reconstructed vascular model. This methodology results in a comprehensive computer-aided diagnostic technique covering the segmentation, reconstruction, and simulation processes. The efficacy of this approach was verified through its application to two cases of aortic arteries and two cases of lower-limb arteries. The accuracy and efficiency of the proposed method were validated by comparison with models reconstructed using 3D Slicer software as the ground truth. The results underscore the method’s capacity to autonomously achieve precision in vascular modeling and promptly execute numerical simulation analyses of blood flow. Additionally, the method yields easily interpretable hemodynamic indicators that can serve as valuable diagnostic aids for medical professionals.

Methods

Figure 1 outlines the comprehensive process employed in this study. The process commences with importing the patient’s CT images, and applying a CNN algorithm to delineate the ROIs in the preprocessed CT images. Subsequently, the 3D reconstruction of the segmented results is facilitated by applying the MC algorithm to generate patient-specific geometric representations. These geometric models are then subjected to hemodynamic simulations using the OpenFOAM solver with predefined parameters. Each procedural step is described in detail in the sections below. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

Figure 1 Flow chart of our proposed method. CT, computed tomography.

Image segmentation

Pre-processing

It is important to preprocess the original CT data before performing deep learning-based image segmentation. A major challenge in medical image segmentation relates to the differences in contrast between the ROI and surrounding organ tissues. CT values, also known as Hounsfield Units (HUs), indicate the extent of X-ray absorption by human tissues, with values in specific ranges corresponding to different organs. Therefore, adjusting CT values can increase the contrast between the ROI and surrounding organ tissues.

CNN

The UNet model, initially proposed by Ronneberger et al. (18) in 2015, has become the predominant architecture for medical image segmentation using CNNs. This model has demonstrated significant success and has served as the backbone for numerous improved networks. In 2021, Dosovitskiy et al. (19) introduced the Vision Transformer (ViT) that incorporated a self-attention mechanism to establish global features in computer vision. Building on this advancement, Chen et al. (20) incorporated the ViT into the UNet, creating the TransUNet model. This integration aimed to address the limitations of CNNs; that is, their confinement to local attention and their inability to capture global relationships.

Despite these advancements, current models encounter substantial challenges in acquiring richer multi-scale features while excluding irrelevant structures and background information due to the considerable size and shape variations among various tissues and organs in medical images. To address this issue, Gao et al. (21) proposed the Res2Net, a new backbone network that builds on the ResNet, and seeks to enhance the network’s capacity to extract multi-scale image features. In the Res2Net framework, a hierarchical residual connection is structured in a single residual block, resembling a residual connection with varying levels. Nevertheless, as the residual block continues to employ convolutional kernels of a fixed size, it fails to overcome the fixed-receptive field limitation.

This study presents the Res2Net-ConvFormer-Dilation-UNet (Res2-CD-UNet) (Figure 2), a multi-scale feature extraction network for medical image segmentation. The Res2Net serves as the backbone of the presented model, with one layer employing dilated convolutions to reduce the number of training parameters and increase the receptive field, thereby enhancing the multi-scale feature extraction capability of the entire backbone network. The ConvFormer (22) is also introduced to the bottom of the encoder to perform global feature extraction on the deep image features. To mitigate the introduction of excessive irrelevant background information through the direct concatenation of low-level features with high-level features in the skip connections, a novel channel feature fusion block is integrated into the skip connection. This block effectively uses the spatial information of low-level features to reduce the impact of background noise and enhance the feature learning ability of the model.

Figure 2 The architecture of the Res2-CD-UNet. ConvFormer, CNN-style transformer; Conv, convolution layers; ReLU, rectified linear unit; Res2-CD-UNet, Res2Net-ConvFormer-Dilation-UNet.

3D reconstruction

This study used VTK for the 3D reconstruction, which includes various image processing and graphics generation algorithms, such as the MC algorithm (23). The MC algorithm is widely used to reconstruct 3D structures from discrete data. It approximates isosurfaces in 3D discrete data fields through linear interpolation. In the context of 3D reconstruction in medical imaging, isosurfaces are defined by setting a specific threshold. At the core of this algorithm are voxels, each representing the smallest unit in the 3D data field, and characterized by eight vertices. These combinations can be systematically grouped into 15 fundamental configurations (24,25) (Figure 3), based on rotational and mapping invariance properties.

Figure 3 Basic constructions of the marching cubes algorithm.

The 3D models generated by the MC algorithm employ multiple triangular facets, which can result in visible stair-step patterns on the model surface. Consequently, surface smoothing is required. In the VTK framework, Laplacian smoothing is used to address this issue. While upholding the topological integrity, this procedure facilitates geometric refinements by repositioning each point to the average of its neighboring points’ coordinates. Through the iterative application of this process, a refined and smoother mesh structure is achieved.

Hemodynamic simulation

In the approach proposed in this article, we used OpenFOAM, a versatile and open-source CFD software package, to streamline converting models obtained from the 3D reconstruction algorithm into OpenFOAM for subsequent meshing and simulation. The mesh generation is performed using the SnappyHexMesh tool in OpenFOAM. Before conducting the CFD simulations, the mesh is checked using the checkMesh function in OpenFOAM to ensure it is suitable for computation. This enables the automatic generation of results on the completion of the hemodynamic simulation.

Experiments

This study introduces a method for automatic vascular segmentation and hemodynamic study using deep learning–based image segmentation algorithms and OpenFOAM software. This study aims to lessen the burden on doctors while enhancing modeling accuracy. In this section, the experimental setup, including the dataset (which consists of a public dataset and a private dataset), evaluation metrics, and predefined parameters, are first outlined. Our proposed medical image segmentation algorithm is then compared with other existing algorithms. Finally, the accuracy of our method is validated by comparing the CFD results from the algorithm-reconstructed (AR) model with that reconstructed using commercial software.

Datasets

The lower-limb artery dataset: this dataset comprises eight lower-limb CT scans from Shanghai Children’s Medical Center affiliated to Shanghai Jiao Tong University. Six scans were selected to train the image segmentation Res2-CD-UNet, while the remaining two were used for testing.

The aortic-artery dataset: the aortic-artery dataset was sourced from the Seg.A.2023 (26) public dataset. The dataset includes 56 annotated CT scans of complete aortic vascular trees. This study focused solely on data from Dongyang Hospital. The dataset was divided into a training set (comprising 12 scans) and a test set (comprising 6 scans).

During the training process, a batch size of 16 and an initial learning rate of 0.001 were applied, and the model underwent training for 30 epochs. Moreover, during the training process, five-fold cross-validation was used to reduce the potential overfitting of the network model.

Experimental setup and metrics

To test the proposed method, the experiments were executed on a Windows 11 platform, using Python (version 3.9), the PyTorch framework, CUDA (version 12.0), a NVIDIA 4090 GPU, VTK (version 9.2.6), and OpenFOAM-10.

The Dice similarity coefficient (DSC) (27) was used to evaluate the segmentation accuracy of the models. The DSC quantifies the similarity between the segmented image and the ground truth by comparing the intersection of the segmented region and the ground truth to the total number of elements in each. A DSC value closer to 1 indicates a higher similarity between the segmented result and the ground truth. The DSC is a crucial tool for evaluating the accuracy of image segmentation algorithms. The DSC is defined as:

Dice(A,B)=2|A∩B||A|+|B|[1]

where A and B represent the prediction and the ground truth, respectively, |A∩B| represent the intersection of the prediction and the ground truth, and |A| and |B| denote the respective number of elements.

To assess the accuracy of the reconstructed models, the relative geometrical error ε was used to compare the inlet and outlet areas, surface, and volume between the AR models and the software-reconstructed (SR) models. The reconstruction error was calculated using the following formula:

ε=Aar−AsrAsr×100%[2]

where A_ar is the parameter of the AR model, and A_sr is the parameter of the SR model.

Predefined parameters

CT values

The CT values, also known as HUs, indicate the extent to which X-rays are absorbed by the tissues. Figure 4 displays the changes in CT images at different CT values. In the original image (239–2,525 HU), it is difficult to accurately outline the vessel contours due to the low contrast between the vessels (highlighted by the red arrows) and the surrounding muscle tissue. Among various preset CT values, the bone window (250–1,000 HU) reduces the influence of the muscle tissue but still has low contrast. The lung window (–450 to 1,500 HU) enhances the brightness of the muscle tissue, worsening the effect. The mediastinal window (30–250 HU) and soft-tissue windows (40–300 HU) increase the contrast between the vessels and muscle tissue, but they excessively highlight the muscle tissue. However, in this study, based on the authors’ experience and visual assessment, the contrast between the blood vessels and muscle tissues improved at the selected specific CT values (161–1,731 HU).

Figure 4 Schematics of CT images at different CT values. The red arrows indicate the regions of interest. CT, computed tomography.

In addition, the CT images were converted from Digital Imaging and Communications in Medicine format to the commonly used Portable Network Graphic format. It is important to note that in medical images, a large part of the image is occupied by a black background with no information. While the vessel region occupies only a tiny part of the image. Excessive irrelevant regions can make it difficult for neural networks to extract features effectively. Thus, the original images need to be cropped properly. Based on the computational performance and efficiency of the networks, this study’s image size was cropped to 224 pixels × 224 pixels.

Iteration of smoothing

In the reconstruction process, choosing the right smoothing level is essential. Excessive iterations can cause the surface of the 3D model to shrink toward the center. Using the SR 3D model as a reference, the changes in the model at different iteration counts were observed (Figure 5). In Figure 5A, the model surface looked jagged due to insufficient iterations. However, with more iterations (Figure 5B), the model surface gradually shrunk toward the center, becoming narrower than the reference. To improve the realism and smoothness of the model surface, this study found that the optimal iteration count was 250. The final 3D reconstructed model was saved in stereolithography format.

Figure 5 Comparison of different smoothing iterations for the same model.

Grid resolution

The grid resolution of the computational domain significantly affects the accuracy and reliability of CFD simulations. In the present study, a comprehensive grid independence assessment was conducted to systematically compare the results obtained from the following three distinct grid levels: a coarse grid with 454,603 nodes; a medium grid with 786,571 nodes; and a fine grid with 1,081,078 nodes. This comparison sought to evaluate the mass flow rate at the outlet and averaged WSS while maintaining consistent boundary conditions across all grid configurations.

The grid independence validation results provide detailed insights into the effect of grid resolution on the accuracy and convergence of the CFD simulations (Table 1). After careful consideration of these results, accounting for both the accuracy of the solutions and the computational time required, the medium-sized grid (786,571 nodes) was selected as the most appropriate grid resolution for subsequent simulation.

Results

2D segmentation

The accuracy of the proposed Res2-CD-UNet for segmentation was evaluated by comparing it with some commonly used networks. As Table 2 shows, the Res2-CD-UNet achieved the best results in both the lower-limb artery and aortic-artery datasets. In the aortic-artery dataset, it achieved an accuracy of 92.76%, which was 1.32% higher than that of the second-best model. In the lower-limb artery dataset, it achieved an accuracy of 94.57%, surpassing the second-best model by 1.12%. However, the Res2-CD-UNet had the highest number of trainable parameters among all the models, reaching 111,554,993.

3D reconstruction

The accuracy of the model reconstruction using the MC algorithm proposed in this study was assessed. Figure 6 displays the aortic arteries and lower-limb arteries reconstructed using the MC algorithm and 3D Slicer for the subjective visual comparison. No significant difference was found between the models reconstructed by the MC algorithm and the commercial software.

Figure 6 Subjective visual comparison of the reconstructed models. (A) Aortic arteries; (B) lower-limb arteries. AR, algorithm-reconstructed; SR, software-reconstructed.

Additionally, to undertake a more quantitative evaluation of the accuracy of the 3D model reconstruction using the MC algorithm, the relative geometrical error ε of the inlet and outlet, the surface, and the volume were calculated (Table 3). The two models closely resembled each other in various key aspects, but the differences were relatively small. The maximum error was only about 2.05% for the lower-limb arteries. These differences did not affect the overall model significantly, indicating that the two models shared a high degree of similarity in terms of their geometric shape and structure.

To further validate the precision of the 3D reconstructed models, several cross-sectional areas perpendicular to the pipe centerline from both sets of the 3D models were extracted. As Figure 7 shows, while the two showed some differences, the values fell within the acceptable error range.

Figure 7 Comparison of cross-sectional areas in the 3D models. (A) Aortic model; (B) lower-limb artery model. AR, algorithm-reconstructed; SR, software-reconstructed.

CFD results

The reconstructed model was then exported to OpenFOAM for the hemodynamic analysis. To evaluate the effect of the 3D reconstruction errors on vascular hemodynamics, a comparative steady CFD analysis was performed on both the SR and the AR models. The analyzed models were subjected to hemodynamic assessment under consistent grid parameters and boundary conditions. The blood was modeled as an incompressible Newtonian fluid with a density of 1,050 kg/m³ and a viscosity of 0.0035 Pa·s, while the wall of the blood vessel was considered impermeable and rigid. The simulation assumed laminar flow for the blood circulation, and the inlet velocities were set to 0.3 m/s for the lower-limb arteries and 0.5 m/s for the aortic arteries. The outlet was configured as a pressure outlet with an outlet pressure of 0 Pa, and the wall was modeled as a no-slip boundary.

The simulations were conducted using the steady-state incompressible simpleFoam solver in OpenFOAM. The qualitative comparison of the CFD results obtained from the models reconstructed using the different methods is visualized in Figures 8,9, which show the contours of WSS and internal streamlines for the aortic and lower-limb arteries, respectively. On close examination of the figures, no substantial disparities in the WSS contours on the surfaces of the models reconstructed using different algorithms were discernible. Additionally, the streamlines delineated consistent vortex regions in corresponding locations for both models.

Figure 8 Contours of wall shear stress and streamlines of the aortic arteries. (A) AR; (B) SR. WSS, wall shear stress; AR, algorithm-reconstructed; SR, software-reconstructed.

Figure 9 Contours of wall shear stress and streamlines of the lower-limb arteries. (A) AR; (B) SR. WSS, wall shear stress; AR, algorithm-reconstructed; SR, software-reconstructed.

Figure 10 presents the results of the quantitative analysis of WSS, showing a high level of agreement between the results derived from the two distinct models. This congruence underscores the robustness and credibility of our employed methodology.

Figure 10 Wall shear stress along the centerline. (A) Aortic arteries; (B) lower-limb arteries. AR, algorithm-reconstructed; SR, software-reconstructed; WSS, wall shear stress.

Discussion

In this study, we compared the Res2-CD-UNet model to other models, including the UNet, AttenUNet, and TransUNet. As detailed in Table 2, our network showed significantly superior performance to that of the exclusively CNN-based networks. This enhancement can be attributed to mitigating constraints inherent in convolutional layers, which exhibit restricted local perspectives and struggle to encompass global contextual information. By amalgamating the ViT into the CNN architecture, our network successfully established extensive global contextual connections, culminating in heightened precision.

Our network uses the enhanced Res2Net block to extract multi-scale features from medical images. Additionally, it integrates a channel feature fusion module in the skip connections to minimize the effect of background noise in the image. These improvements were instrumental in achieving the highest accuracy results.

The objective of this study was to enable rapid and automatic hemodynamic analysis, emphasizing the significance of accurate reconstructed models. Consequently, we compared the hemodynamic results of the AR models with those of the SR models. The results revealed strong agreement between the reconstructed models using different methods, which validated the proposed approach. Notably, the overall computational time for the process significantly decreased from several hours to a few minutes, which would substantially enhance diagnostic efficiency in clinical settings.

Conclusions

This study introduced a novel approach for examining vascular hemodynamics by applying deep learning-based medical image segmentation, vessel reconstruction, and OpenFOAM simulation. The developed method facilitates the automated segmentation, 3D reconstruction, and CFD simulation of arterial regions in CT images. The comparison of the AR models with those generated using the open-source software 3D Slicer for flow-field analysis revealed a high concordance in the hemodynamic analysis between the two methods. These findings suggest that the proposed method exhibits high accuracy and enables the rapid and intuitive visualization of hemodynamic changes in the arteries, and thus could assist healthcare professionals to formulate precise treatment strategies for patients. Notably, there was a substantial reduction in the overall computational time required for the process, from multiple hours to a few minutes, which represents a significant increase in diagnostic efficiency and throughput. Although deep-learning CNNs can quickly segment ROIs, they inevitably increase the demand placed on computational resources. Additionally, the efficient post-processing and extraction of hemodynamic parameters remains a challenge that needs to be addressed. Therefore, in our future research, we intend to focus on the lightweight optimization of the network model and the simplification of the post-processing steps.

Acknowledgments

None.

Footnote

Funding: This study received funding from the National Natural Science Foundation of China (No. 12172227).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1721/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work, including ensuring that any questions related to the accuracy or integrity of any part of the work have been appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Naser MA, Majeed AA, Alsabah M, Al-Shaikhli TR, Kaky KM. A review of machine learning’s role in cardiovascular disease prediction: recent advances and future challenges. Algorithms 2024;17:78.
2. Ahmed H, Younis EMG, Hendawi A, Ali AA. Heart disease identification from patients’ social posts, machine learning solution on spark. Future Gener Comput Syst 2020;111:714-22.
3. Chen J, Lu XY. Numerical investigation of the non-Newtonian pulsatile blood flow in a bifurcation model with a non-planar branch. J Biomech 2006;39:818-32. [Crossref] [PubMed]
4. Zhang Y, Lu Q, Feng J, Yu P, Zhang S, Teng Z, Gillard JH, Song R, Jing Z. A pilot study exploring the mechanisms involved in the longitudinal propagation of acute aortic dissection through computational fluid dynamic analysis. Cardiology 2014;128:220-5. [Crossref] [PubMed]
5. Condemi F, Serra R, De Franciscis S, Fragomeni G. Popliteal artery aneurysm: a bioengineering study on pathogenesis. In: Longo F, Frascio M, Novak V, Rozenblit JW, Merkuryev Y, Bruzzone AG, editors. IWISH 2015: 4th International Workshop on Innovative Simulation for Health Care; 2015 Sep 21-23; Bergeggi, Italy. P. 10-6.
6. Hyde-Linaker G, Barrientos PH, Stoumpos S, Kingsmore DB, Kazakidi A. Patient-specific computational haemodynamics associated with the surgical creation of an arteriovenous fistula. Med Eng Phys 2022;105:103814. [Crossref] [PubMed]
7. Soliveri L, Bozzetto M, Brambilla P, Caroli A, Remuzzi A. Hemodynamics in AVF over time: A protective role of vascular remodeling toward flow stabilization. Int J Artif Organs 2023;46:547-54. [Crossref] [PubMed]
8. Abhilash HN, Yanagita Y, Pai R, Zuber M, Tamagawa M. K P, Kamath S G, R P, Barboza ABV, Rao VRK, Khader SMA. Effect of vascular geometry on haemodynamic changes in a carotid artery bifurcation using numerical simulation. Clin Neurol Neurosurg 2024;237:108153. [Crossref] [PubMed]
9. Kang T, Mukherjee D, Kim JM, Park KY, Ryu J. Effects of progressive carotid stenosis on cerebral haemodynamics: aortic-cerebral 3D patient-specific simulation. Eng Appl Comput Fluid Mech 2021;15:830-47.
10. Caddy HT, Kelsey LJ, Parker LP, Green DJ, Doyle BJ. Modelling large scale artery haemodynamics from the heart to the eye in response to simulated microgravity. NPJ Microgravity 2024;10:7. [Crossref] [PubMed]
11. Luan J, Qiao Y, Mao L, Fan J, Zhu T, Luo K. The role of aorta distal to stent in the occurrence of distal stent graft-induced new entry tear: A computational fluid dynamics and morphological study. Comput Biol Med 2023;166:107554. [Crossref] [PubMed]
12. Colombo M, Bologna M, Garbey M, Berceli S, He Y, Rodriguez Matas JF, Migliavacca F, Chiastra C. Computing patient-specific hemodynamics in stented femoral artery models obtained from computed tomography using a validated 3D reconstruction method. Med Eng Phys 2020;75:23-35. [Crossref] [PubMed]
13. Cao L, Shi R, Ge Y, Xing L, Zuo P, Jia Y, Liu J, He Y, Wang X, Luan S, Chai X, Guo W. Fully automatic segmentation of type B aortic dissection from CTA images enabled by deep learning. Eur J Radiol 2019;121:108713. [Crossref] [PubMed]
14. Duan X, Li Y, Wang J. Coronary artery CTA image segmentation and three-dimensional visualization based on U-Net. Proceedings of the 2020 International Symposium on Artificial Intelligence in Medical Sciences. ISAIMS 2020:64-68.
15. Chen Y, Sun P. The research and practice of medical image 3D reconstruction platform. In: Proceedings 2014 IEEE International Conference on Security, Pattern Analysis, and Cybernetics (SPAC); Wuhan, China, 2014. P. 305-8.
16. Yin J, Yang G. Remote three-dimensional reconstruction for medical images based on Browser/Server framework. In: 2016 9th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI); Datong, China, 2016. P. 1877-81.
17. Mittal R, Malik V, Singla G, Kaur A, Singh M, Mittal A. 3D reconstruction of brain tumors from 2D MRI scans: An improved marching cube algorithm. Biomed Signal Process Control 2024;91:105901.
18. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. MICCAI 2015: Medical Image Computing and Computer-Assisted Intervention; 2015, p. 234-41.
19. Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai XH, Unterthiner T, Dehghani M, Minderer M, Heigold G, Gelly S, Uszkoreit J, Houlsby N. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. Preprint at https://doi.org/10.48550/arXiv.2010.11929 (2021).
20. ChenJNLuYYYuQHLuoXDAdeliEWangYLuLYuilleALZhouYY. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation. Preprint at http://arxiv.org/abs/2102.04306 (2021).
21. Gao SH, Cheng MM, Zhao K, Zhang XY, Yang MH, Torr P. Res2Net: A New Multi-Scale Backbone Architecture. IEEE Trans Pattern Anal Mach Intell 2021;43:652-62. [Crossref] [PubMed]
22. Lin X, Yan ZQ, Deng XB, Zheng CS, Yu L. ConvFormer: plug-and-play CNN-style transformers for improving medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2023: 26th International Conference, Vancouver, BC, Canada, October 8–12, 2023, Proceedings, Part IV. Springer-Verlag, Berlin, Heidelberg, 642-51.
23. Lorensen WE, Cline HE. Marching cubes: a high resolution 3D surface construction algorithm. Seminal graphics: pioneering efforts that shaped the field, Volume 1. Association for Computing Machinery, New York, NY, USA, 347-353. ACM SIGGRAPH Computer Graphics 1987;21:163-9.
24. Guo C, Zhuang Z, Josephraj AN. 3D reconstruction of ultrasonic carotid artery based on fractal dimension and marching cubes. In: Tenth International Conference on Digital Image Processing (ICDIP 2018); SPIE, Shanghai, China, 2018.
25. Visutsak P. Marching Cubes and Histogram Pyramids for 3D Medical Visualization. J Imaging 2020;6:88. [Crossref] [PubMed]
26. Radl L, Jin Y, Pepe A, Li J, Gsaxner C, Zhao FH, Egger J. AVT: Multicenter aortic vessel tree CTA dataset collection with ground truth segmentation masks. Data Brief 2022;40:107801. [Crossref] [PubMed]
27. Liu T, Lu YH, Zhang Y, Hu JH, Gao C. A bone segmentation method based on Multi-scale features fuse U2Net and improved dice loss in CT image process. Biomed Signal Process Control 2022;77:103813.
28. OktayOSchlemperJLe FolgocLLeeMHeinrichMMisawaKMoriKMcdonaghSHammerlaNYKainzBGlockerBRueckertD. Attention U-Net: learning where to look for the pancreas. Preprint at http://arxiv.org/abs/1804.03999 (2018).