Automated identification, hemodynamic visualisation and quantification of abdominal aortic aneurysms with 4D flow magnetic resonance imaging

Introduction

An abdominal aortic aneurysm (AAA) is a pathological dilatation of the aorta in the abdominal region. Rupture of an AAA is associated with high mortality rates above 80% (1). To prevent rupture, elective surgery can be considered when the AAA reaches a certain diameter threshold of 5 cm for women and 5.5 cm for men (1). However, this diameter threshold alone is not specific enough to prevent rupture in all patients (2,3). Quantitative magnetic resonance imaging (MRI) may improve rupture risk assessment by providing additional biomarkers assessing the physiology of the aneurysm next to structural information alone.

Four dimensional (4D) flow MRI [time-resolved, three dimensional (3D) phase contrast imaging with three-directional velocity encoding (VENC)] allows for assessment and visualisation of hemodynamics and related derived parameters over the cardiac cycle. One hemodynamic biomarker of particular interest is wall shear stress (WSS), as regions with low WSS have been associated with aneurysm formation and rupture (4-6). Currently, 4D flow MRI is mostly utilised in a research setting whereas adoption in the clinical workflow remains limited. According to both a Delphi analysis involving 18 experts in the field and the updated 4D flow cardiovascular consensus statement, one of the limitations that hampers widespread adoption and application of 4D flow MRI in the clinic is cumbersome and time-consuming data processing (7,8). To address this limitation the authors suggested efficient image processing with minimal user dependence. For these reasons, there has been interest in developing automated post-processing pipelines for 4D flow MRI.

An important part of automated post-processing is automated vessel segmentation. Typically, a phase-contrast magnetic resonance angiography (PC-MRA) is generated from the 4D flow phase and magnitude data for segmentation (9-13). Automated segmentation of the thoracic aorta and pulmonary artery based on PC-MRA has been demonstrated with both deep learning- and atlas-based methods (10-13). However, AAA visibility on PC-MRA is often limited because of low flow velocity within the aneurysm, which makes manual and automated segmentation based on PC-MRA in AAA patients more challenging. 3D cine balanced steady state free precession (bSSFP) MRI has better structural contrast compared to the PC-MRA and enables segmentation of both lumen and intraluminal thrombus (ILT) which could provide additional information about AAA progression (14).

Next to automated segmentation, research has also focused on automating other parts of 4D flow MRI post-processing. For example, an open-source Python-based pipeline was created for loading and visualising 4D flow MRI data from patients with pulmonary hypertension (15). Another automated tool was developed for cranial 4D flow MRI including interactive vessel selection and hemodynamic quantification (16,17). Also, for cardiac 4D flow MRI a pipeline was designed using deep learning for biventricular segmentation (18). Lastly, a cardiac tool was developed including full integration of the 4D flow analysis on the scanner involving both flow visualisation and quantification (13). Despite all these efforts, automated post-processing tools for 4D flow MRI in the abdomen and more specifically in AAA are lacking.

Therefore, the aim of the current study is to develop an automated post-processing pipeline for 4D flow MRI in AAA patients, incorporating aneurysm identification, via automated segmentation of 3D cine bSSFP MRI, blood flow visualisation and hemodynamic parameter quantification.

Methods

Study design and population

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The Amsterdam UMC institutional ethics committee approved the study within the scope of the Dutch Medical Research Involving Human Subjects Act (WMO) (No. NL80822.029.22). Written informed consent was obtained from all participants. Patients with an asymptomatic AAA measuring at least 30 mm in diameter (1), based on computed tomography (CT), ultrasound (US) or MRI measurements, were recruited for this study, which is part of a single-centre investigation (ClinicalTrials.gov ID: NCT05976711). Patients were excluded from study participation in case of severely reduced renal function [estimated glomerular filtration rate (eGFR) <20 mL/min/1.73 m²], supra- or pararenal AAA, previous AAA repair, cardiac arrhythmias not currently under treatment, inflammatory, infectious, or mycotic AAA, vasculitis, or a connective tissue disorder.

MRI acquisition

All MRI scans were performed on a 3.0 T Philips MR7700 MRI system equipped with software release 5.9 (Philips Medical Systems, Best, The Netherlands), a dStream 16-channel Torso coil and a 12-channel posterior coil. Participants underwent an abdominal MRI scan including free-breathing 4D flow MRI and 3D cine bSSFP MRI (19). As part of the comprehensive scan protocol patients were injected with a Gadolinium contrast agent (Dotarem, Guerbet, Villepinte, France) for dynamic contrast enhanced MRI, approximately 15 and 25 minutes prior to 4D flow and 3D cine bSSFP acquisition, respectively. Electrocardiography (ECG) signals were acquired to allow for retrospective cardiac binning. In case of unstable ECG signal peripheral pulse signal was used instead. The sagittal field of view (FOV) was positioned to encompass the region from the renal arteries to the aortic bifurcation into the iliac arteries. The acquisition parameters for 4D flow MRI were: acquired and reconstructed spatial resolution (1.6 mm)³ and (1.0 mm)³, respectively; temporal resolution ~40 ms (24 cardiac phases); echo time (TE) =2.78 ms; repetition time (TR) =4.58 ms; VENC 52 cm/s in all three velocity directions and scan duration between 7 and 11 minutes depending on the size of the patient. The selected VENC was based on previous research in AAA patients (9). A two dimensional (2D) phase contrast scout scan was placed below the renal arteries to detect potential aliasing occurring with the chosen VENC. In two patients the VENC was increased to 75 cm/s because of observed aliasing in the 2D phase contrast scout scan. 3D cine bSSFP MRI was acquired with the same spatial resolution and FOV as the 4D flow MRI. It was reconstructed to a temporal resolution of 67 ms (15 cardiac phases) further acquisition parameters were: TE =1.44 ms; TR =2.89 ms; flip angle =40° and scan duration 4–6 minutes depending on FOV tailored to patient’s size. A detailed overview of the 3D cine bSSFP MRI acquisition and reconstruction is described elsewhere (19,20). In short, prospective undersampling in multiple dimensions (PROUD) (21) was used to sample a variable density pseudo-spiral Cartesian pattern. Respiratory gating was performed based on an in-bore camera (VitalEye) tracking the breathing position. Consequently, data was sorted in a five dimensional (5D) dataset with 15 cardiac and 4 respiratory phases. First, images were averaged over the cardiac dimension followed by 3D non-rigid registration to the end-expiration phase. Resulting respiratory displacements were applied to the 5D images to create respiratory motion-compensated data.

Segmentation

Figure 1 displays both a PC-MRA and 3D cine bSSFP MRI of the same patient, highlighting the superior structural contrast achieved with 3D cine bSSFP MRI. To enable quick and automated segmentation of 3D cine bSSFP data, an nnU-Net (22) was trained to segment the aorta (lumen and ILT separately). To train the network manual segmentations were created using Mimics (version 23.0, Materialise, Leuven, Belgium, RRID:SCR_012153) by a technical physician (E.A.) with 3 years of experience in segmenting vascular structures. All efferent vessels of the abdominal aorta were not included (i.e., renal arteries, superior mesenteric artery and inferior mesenteric artery) in the segmentation. After bifurcation of the aorta into the iliac arteries, further bifurcations of the iliac arteries (externa, interna) were included in the segmentation when present within the FOV. Sixteen baseline scan segmentations of which twelve contained ILT, were used for training the nnU-Net with 5-fold cross-validation, in 3D full resolution. In 5-fold cross-validation, the dataset is split into five folds. For each run, one fold is used for validation and the remaining four folds for training, resulting in five trained models. The resulting nnU-Net segmentations were generated using an ensemble of the five models. To enable fair validation, the split of training and validation data was adjusted to ensure that the validation split would contain at least one patient with ILT. Dice-similarity coefficients (DSCs) were calculated to quantify the spatial overlap between the ground truth and the segmentations generated by the nnU-Net for lumen and ILT segmentations separately. A DSC of 1 represents perfect spatial overlap, whereas a DSC of 0 indicates no spatial overlap. The DSC is used by the nnU-Net as a loss function and evaluation metric to train the model. To allow for more robust analysis of the nnU-Net performance, the 95% percentile Hausdorff distance (HD95) was calculated as well (separately for lumen and ILT), which quantifies the spatial discrepancy (boundary discrepancy) between predicted and ground truth segmentations. This metric is a non-negative real number given in millimeters, whereby a value of 0 mm indicates perfect alignment.

Figure 1 Visualisation of lumen-background contrast differences on PC-MRA and 3D cine bSSFP. (A) PC-MRA generated based on 4D flow MRI data. The magnified box shows poor contrast between aneurysm lumen and background. (B) 3D cine bSSFP scan from the same patient which provides good contrast between lumen, thrombus and background. 3D, three dimensional; 4D, four dimensional; bSSFP, balanced steady state free precession; MRI, magnetic resonance imaging; PC-MRA, phase-contrast magnetic resonance angiography.

Automated post-processing pipeline

The automated 4D flow MRI post-processing pipeline was developed in MATLAB (version R2021a, MathWorks, Natrick, MA, USA, RRID:SCR_001622) utilising features from existing open-source 4D flow processing software (23) and is available on GitHub (https://github.com/schrau24/AAA_4Dflow_auto) (24). A schematic overview of the automated post-processing pipeline is given in Figure 2. Quantification of parameters is depicted as a parallel box to the post-processing box. No interaction with the user is needed. Both a segmentation of the lumen alone and a segmentation of the lumen and ILT combined, should be available to the pipeline. In the first step of the post-processing pipeline, all input and segmented data were cropped to reduce computing time. Cropping was based on the extent of the combined segmentation with a margin of ten voxels. Cropped 4D flow and 3D cine bSSFP data served as input for the second block of post-processing. First a multimodal linear registration was done to register the 4D flow data on the 3D cine bSSFP data, in case the patient moved between the subsequent scans. Laplacian unwrapping was applied to remove potential aliasing caused by high velocities that exceeded the chosen VENC (25). Lastly divergence-free denoising was applied to reduce noise in the velocity data while preserving flow structures and adhering to the constraint of incompressible flow (26). The output of this second block is used to calculate and visualise the flow velocity over all time frames. Centerline extraction was performed using the MATLAB internal function “bwskel” which thins the lumen, in an iterative process, to a one-voxel thick representation. In cases with ILT, the centerline is still defined by the lumen, since alignment with the blood flow direction is required for accurate cross-sectional velocity calculations (not utilised in current analysis). One centerline was created for the proximal aorta and aneurysm. After bifurcation of the aorta two extra centerlines were formed for the iliac arteries. In each orthogonal cross-section along the centerline, ellipse fitting was used to determine the diameter of the segmented lumen along the aortic centerline and one of the iliac centerlines (27). Modifications of the resulting diameter graphs were necessitated due to tortuous anatomical variations of the aorta, which are detailed in Figure S1. A 3 cm threshold (1) was applied on the adjusted diameter graphs to extract the aneurysm from the aortic volume. The extracted aneurysm volume combined with the cropped segmentation data and pre-processed 4D flow MRI data served as input for WSS calculations within the aneurysm. In literature regional variations of WSS are described in complex vascular geometry (28-30). Also, AAA rupture locations have been associated with regional low WSS (4). To enable region-specific assessment of WSS, the extracted aneurysm volume was divided into eight quadrants, defined by anterior, posterior, left and right directions each split into upper and lower parts. The entire pipeline was run twice for each patient, once using the manual segmentation and once using the nnU-Net segmentation as input. The volume of the resulting extracted aneurysms was compared as well as the mean and peak WSS values estimated based on those volumes.

Figure 2 Overview of the automated AAA post-processing pipeline. The input data is depicted in green. Segmented data input (in this case based on 3D cine bSSFP data) are depicted in orange. Both serve as input for the post-processing pipeline depicted in turquoise with quantification steps shown in the pink box. 3D, three dimensional; 4D, four dimensional; AAA, abdominal aortic aneurysm; bSSFP, balanced steady state free precession; Seg., segmentation; WSS, wall shear stress.

WSS estimation

The method for quantifying 3D WSS within the aneurysm using 4D flow MRI has been reported elsewhere (31,32), with its reproducibility previously established (test-retest voxel-by-voxel Bland-Altman analysis: mean difference 0.01 Pa and limits of agreement 0.23 Pa) (9). In summary, the WSS was computed for each time frame using a tensor-based approach. The shear stress tensor was calculated by multiplying the nominal blood viscosity value 0.0032 Pa·s, the rate of deformation tensor and the normal vector perpendicular to the lumen wall. To simplify the calculations, the aortic geometry was temporarily rotated to align the z-axis to the normal vector of the aortic wall. At each point along the vessel wall, velocity gradients were approximated by fitting a spline through three interpolated velocity values along the inward normal. The resulting WSS vectors were calculated after an inverse rotation back to the original coordinate system. Both the mean and peak WSS values were calculated for every time frame over all voxels along the aneurysm wall. The peak WSS per time frame was calculated by using the 95th percentile of the values to prevent outliers from corrupting the data. From these mean and peak WSS values per time frame, the time frames with the minimum and maximum values were selected. Resulting in four variables: minimum mean WSS, maximum mean WSS, minimum peak WSS and maximum peak WSS.

Statistical analysis

Continuous parameters are presented as mean ± standard deviation (SD) for normally distributed data, or as median with interquartile range (IQR) for non-normally distributed data. The Shapiro-Wilk test was applied to test for normality of the acquired WSS data. The Pearson correlation coefficients were calculated to assess potential correlations between the mean and peak WSS values calculated based on nnU-Net derived segmentations and manual segmentations. Bland-Altman plots were generated to visually assess the agreement between the WSS values originating from the two segmentation methods.

Statistical significance was set at a two-tailed P value of 0.05. R software (R Core Team, version 1.4.1717, Vienna, Austria) was used for the statistical analyses.

Results

Resulting post-processing pipeline

Figure 3 shows a visual representation of the most important steps in the resulting post-processing pipeline for 4D flow MRI using one example AAA patient. The pipeline was tested using data from 14 patients, all of which were successfully processed. Two of the 16 available patient datasets were excluded from testing the pipeline: one due to missing 4D flow MRI data, and the other because the aortic diameter did not reach 30 mm. Processing time per patient was approximately 30 minutes using 6 CPU cores and 45GB of RAM.

Figure 3 A visual representation of a few steps within the post-processing pipeline for one AAA patient. (I) First, a centerline is generated based on a thinning algorithm. (II) Along the centerline, ellipse fitting is applied to each cross-section to determine the maximum diameter. (III) Maximum diameter along the centerline is visualised in a graph. Based on a threshold of ≥3 cm the aneurysm is extracted from the segmented volume. (IV) Resulting aneurysm volume including ILT in yellow. (V) Flow velocity is calculated over the entire volume. (VI) WSS is calculated within the aneurysm. (VII, VIII) The aneurysm volume is divided into 8 quadrants. (IX) Peak WSS for all upper quadrants is estimated over all time frames. AAA, abdominal aortic aneurysm; ILT, intraluminal thrombus; WSS, wall shear stress.

Performance of the nnU-Net

The performance of the nnU-Net was excellent for lumen segmentation and good for thrombus segmentation, with DSCs of 0.93 (0.05) and 0.80 (0.26) respectively. The HD95 was 4.87 (8.59) mm for the lumen segmentation and 6.04 (3.71) mm for the thrombus segmentation. In Figure 4, the segmentation evaluation metrics are depicted in a boxplot. There is one outlier with a DSC of 0 in the thrombus segmentation data. This occurred because a small thrombus, missed in the manual segmentation, was correctly identified by the nnU-Net, see Figure 5. In this case, the nnU-Net outperformed the manual segmentation. Next to that, a notable outlier was found for the HD95 in the thrombus segmentation data as well. This outlier was caused by multiple small falsely segmented regions identified by the nnU-Net at the two most caudal slices of the FOV.

Figure 4 Boxplots for Dice similarity coefficients and 95% Hausdorff distances for both lumen (red) and thrombus (blue) segmentations depicted in respectively the upper and lower graph.

Figure 5 Visualisation of a case where the nnU-Net detected a thrombus overlooked in manual segmentation. (A) Axial slice of an aorta with small thrombus. (B) nnU-Net segmentation with detected lumen (purple) and thrombus (pink). (C) Manual segmentation with only lumen (turquoise).

In Figure 6, sagittal and transversal slices of a 3D cine bSSFP scan of two patients are visualised to enable comparison between the nnU-Net based and manual segmentations. The upper row illustrates a case with excellent nnU-Net performance, whereas the lower row depicts a case with relatively low DSC. In the latter, the patient had an extensive, irregularly shaped thrombus that made the manual segmentation challenging. Nevertheless, the nnU-Net segmentation appears accurate.

Figure 6 Sagittal and transversal slices of 3D cine bSSFP MRI in two AAA patients (one row per patient). nnU-Net segmentations are depicted in the second and fifth column containing both lumen (purple) and thrombus (pink). Columns three and six contain the manual segmentation with lumen (turquoise) and thrombus (red). (A-F) A case that was easy to segment yielding high DSCs of 0.96 and 0.89 for lumen and thrombus segmentation, respectively. (G-L) A case with irregular thrombus volume, resulting in lower DSCs of 0.79 and 0.64 for lumen and thrombus, respectively. 3D, three dimensional; AAA, abdominal aortic aneurysm; bSSFP, balanced steady state free precession; DSCs, Dice-similarity coefficients; MRI, magnetic resonance imaging; Seg., segmentation.

The median difference between the calculated aneurysm volume based on the manual and nnU-Net segmentation was 2.28 (4.09) mL. The median differences between the estimated mean and peak WSS values are depicted in Table 1. Significant correlations (ρ =0.92 or higher) were found between WSS values derived using the manual and nnU-Net segmentations.

The timeframes with maximum WSS values showed more variability between the two segmentation methods compared to the timeframes with minimum WSS values, as indicated by the wider limits of agreement of the corresponding Bland-Altman plots in Figure 7.

Figure 7 Bland-Altman plots for the (A) minimum mean WSS, (B) maximum mean WSS, (C) minimum peak WSS, and (D) maximum peak WSS values assessed within the extracted aneurysm based on the manual and nnU-Net segmentation. Diff, difference; max, maximum; min, minimal; SD, standard deviation; WSS, wall shear stress.

Discussion

In this study an automated post-processing pipeline for 4D flow MRI in AAA patients was developed incorporating aneurysm extraction, blood flow visualisation and quantification of hemodynamic parameters. Segmentations were based on 3D cine bSSFP data with the same FOV and spatial resolution as the 4D flow MRI data to facilitate robust segmentation. To accelerate the segmentation, an nnU-Net was trained which demonstrated accurate performance. WSS values derived from nnU-Net based segmentation showed excellent correlation with the estimated WSS values based on manual segmentation in the same patient. The post-processing pipeline was robust for the varying anatomies present in the cohort and aneurysm extraction was successful in all cases. The derived aneurysm volume was divided into eight quadrants to enable regional WSS analysis.

Automated post-processing of 4D flow MRI data has received increasing attention, given its pivotal role in facilitating the clinical implementation of 4D flow MRI (15,16,18). Jin et al. (13) went one step further and developed an automated framework for inline cardiac 4D flow MRI analysis on the scanner with a processing time of approximately five minutes without user interaction. This is a promising advancement to further streamline 4D flow MRI adoption. However, their framework did not incorporate WSS calculations, which represent the most time-consuming part of our pipeline. Moreover, the development of a cardiac tool eliminates the need to address ILT (33,34). In the current study we focus on patients with an AAA. This presents inherent challenges because of tortuous vessel anatomy and the presence of ILT in the majority of patients, which often results in irregular lumen shapes. Nevertheless, post-processing including aneurysm extraction was successful in all cases.

Within this study an nnU-Net was trained to segment the abdominal aorta, AAA and iliac arteries from 3D cine bSSFP images. The main focus of our study was developing a robust post-processing pipeline for 4D flow MRI. However, segmentations are a necessary input for the developed pipeline, and using an nnU-Net to generate these segmentations reduces user dependence (which aligns with the automated character of the pipeline), provides accurate results and saves time. Several studies have focused on segmentation of 4D flow MRI data based on a U-net architecture leveraging datasets considerably larger than the one used in this study (10,35,36). Nevertheless, these studies reported DSCs comparable to those achieved in our analysis. This may be attributed to the use of 3D cine bSSFP MRI instead of the PC-MRA for segmentation in the current study. In a previous study the same acquisition was used to train an nnU-Net for segmentation of the thoracic aorta in healthy volunteers showing DSCs in line with our findings (37). Ideally, more data would be used to train an nnU-Net for segmentation of the AAA. Nevertheless, the results obtained with the model trained in this study are consistent with those reported in literature, suggesting that it can be reliably used.

A strength of the current study is the use of 3D cine bSSFP MRI to enable robust segmentation of the AAA. This sequence provides excellent structural information to identify lumen and thrombus whereas AAA lumen and background are difficult to distinguish on the conventionally used PC-MRA because of slow flow. Even though acquiring a 3D cine bSSFP scan will add 4-6 minutes to the scan protocol, it contributes to more reliable WSS measurements because the estimated WSS values are heavily dependent on accurate lumen boundary segmentation (38,39).

Within our group, the post-processing pipeline has been tested on thoracic and fetal 4D flow MRI data. For these applications small changes were needed and some functionality of the pipeline was redundant (i.e., functions for registration of 3D cine bSSFP and 4D flow MRI data and thrombus volume calculation). However, the implementation was straightforward and robust, which suggests possible adaptation of the pipeline to other anatomical regions.

A limitation of the current study was the small number of AAA cases available to develop and test the pipeline. Therefore, the developed pipeline may encounter errors when applied to anatomies that deviate substantially from those represented in this study. Future work could improve generalisability by including multi-centre and multi-vendor data. In addition, the small sample size meant that splitting the data into training, test, and validation sets for nnU-Net would have resulted in small groups, leading to poor training and unreliable evaluation. Therefore, we opted for 5-fold cross-validation, whereby each case is used for both training and validation, maximising data utilisation and providing a more reliable performance estimate. Nevertheless, this results in the absence of an independent test set to assess model performance.

With the developed post-processing pipeline clinical implementation of 4D flow MRI is one step closer. Nevertheless, there are multiple steps remaining. Next to time-consuming data processing, which may be addressed by incorporating automated post-processing pipelines, single user-selected VENC, long and unpredictable scan times and data storage are still challenges outlined in the 4D flow cardiovascular magnetic resonance consensus statement of 2023 (8). Moreover, if implementation of an optimized 4D flow MRI sequence in the clinical workflow is realised, still the added value of 4D flow MRI in predicting rupture risk in AAA patients remains unknown (40). To validate 4D flow MRI derived biomarkers such as WSS for AAA rupture risk prediction, longitudinal studies are necessitated in the future: first, to assess correlations between 4D flow MRI derived parameters and aortic diameter, the gold standard in aneurysm risk prediction; second, to assess correlations between 4D flow MRI derived parameters and clinical endpoints such as rupture, death or repair.

Conclusions

An automated post-processing pipeline was developed specifically for 4D flow MRI in patients with AAA utilising 3D cine bSSFP MRI for segmentations. The pipeline is robust for varying anatomies caused by tortuosity and ILT and could facilitate 4D flow MRI implementation in the clinical workflow of AAA patients. Next steps should involve testing the proposed pipeline on multi-centre and multi-vendor data and investigating WSS as a biomarker for AAA progression.

Acknowledgments

Figure 2 was created with BioRender.com.

Footnote

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1335/dss

Funding: This work was supported by Health~Holland (No. LSH-TKI 25379).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1335/coif). E.A., W.S., A.N., and K.K.Y. report that the authors received a grant from Health~Holland to do the MARVY project of which this study is part. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by institutional ethics committee of the Amsterdam UMC (No. NL80822.029.22) and written informed consent was obtained from all individual participants.

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. Wanhainen A, Van Herzeele I, Bastos Goncalves F, Bellmunt Montoya S, Berard X, Boyle JR, et al. Editor's Choice -- European Society for Vascular Surgery (ESVS) 2024 Clinical Practice Guidelines on the Management of Abdominal Aorto-Iliac Artery Aneurysms. Eur J Vasc Endovasc Surg 2024;67:192-331. [Crossref] [PubMed]
2. Laine MT, Vänttinen T, Kantonen I, Halmesmäki K, Weselius EM, Laukontaus S, Salenius J, Aho PS, Venermo M. Rupture of Abdominal Aortic Aneurysms in Patients Under Screening Age and Elective Repair Threshold. Eur J Vasc Endovasc Surg 2016;51:511-6. [Crossref] [PubMed]
3. Earnshaw JJ. The Indication for Elective Repair of Abdominal Aortic Aneurysm Should Be Reviewed. Eur J Vasc Endovasc Surg 2021;61:7-8. [Crossref] [PubMed]
4. Boyd AJ, Kuhn DC, Lozowy RJ, Kulbisky GP. Low wall shear stress predominates at sites of abdominal aortic aneurysm rupture. J Vasc Surg 2016;63:1613-9. [Crossref] [PubMed]
5. Trenti C, Ziegler M, Bjarnegård N, Ebbers T, Lindenberger M, Dyverfeldt P. Wall shear stress and relative residence time as potential risk factors for abdominal aortic aneurysms in males: a 4D flow cardiovascular magnetic resonance case-control study. J Cardiovasc Magn Reson 2022;24:18. [Crossref] [PubMed]
6. Bürk J, Blanke P, Stankovic Z, Barker A, Russe M, Geiger J, Frydrychowicz A, Langer M, Markl M. Evaluation of 3D blood flow patterns and wall shear stress in the normal and dilated thoracic aorta using flow-sensitive 4D CMR. J Cardiovasc Magn Reson 2012;14:84. [Crossref] [PubMed]
7. van Schuppen J, van der Hulst AE, den Harder JM, Gottwald LM, van Luijk RD, van den Noort JC, Nelissen JL, Coerkamp CF, Boekholdt SM, Groot PFC, Nederveen A, van Ooij P, Planken RNDelphi Panel Members. Prerequisites for Clinical Implementation of Whole-Heart 4D-Flow MRI: A Delphi Analysis. J Magn Reson Imaging 2025;61:1618-28. [Crossref] [PubMed]
8. Bissell MM, Raimondi F, Ait Ali L, Allen BD, Barker AJ, Bolger A, et al. 4D Flow cardiovascular magnetic resonance consensus statement: 2023 update. J Cardiovasc Magn Reson 2023;25:40. [Crossref] [PubMed]
9. Aalbregt E, Indrakusuma R, Jalalzadeh H, Planken RN, van Schuppen J, Meijboom L, Balm R, Nederveen AJ, Yeung KK, van Ooij P. Four-Dimensional Flow MRI-Derived Hemodynamics in Abdominal Aortic Aneurysms: Reproducibility and Associations With Diameter, Intraluminal Thrombus Volume, and Vorticity. J Magn Reson Imaging 2024;60:878-88. [Crossref] [PubMed]
10. Berhane H, Scott M, Elbaz M, Jarvis K, McCarthy P, Carr J, Malaisrie C, Avery R, Barker AJ, Robinson JD, Rigsby CK, Markl M. Fully automated 3D aortic segmentation of 4D flow MRI for hemodynamic analysis using deep learning. Magn Reson Med 2020;84:2204-18. [Crossref] [PubMed]
11. Fujiwara T, Berhane H, Scott MB, Englund EK, Schäfer M, Fonseca B, Berthusen A, Robinson JD, Rigsby CK, Browne LP, Markl M, Barker AJ. Segmentation of the Aorta and Pulmonary Arteries Based on 4D Flow MRI in the Pediatric Setting Using Fully Automated Multi-Site, Multi-Vendor, and Multi-Label Dense U-Net. J Magn Reson Imaging 2022;55:1666-80. [Crossref] [PubMed]
12. Bustamante M, Petersson S, Eriksson J, Alehagen U, Dyverfeldt P, Carlhäll CJ, Ebbers T. Atlas-based analysis of 4D flow CMR: automated vessel segmentation and flow quantification. J Cardiovasc Magn Reson 2015;17:87. [Crossref] [PubMed]
13. Jin N, Berhane H, Davids R, Pruitt A, Baraboo J, Bi X, Markl M, Chow K. Fully Automated Inline 4D Flow MRI Visualization and Flow Quantification. Journal of Cardiovascular Magnetic Resonance 2024;26:100171.
14. Boyd AJ. Intraluminal thrombus: Innocent bystander or factor in abdominal aortic aneurysm pathogenesis? JVS Vasc Sci 2021;2:159-69. [Crossref] [PubMed]
15. Arbelaiz A, Bóbeda J, García S, Ruíz-Cabello J, García-Cosío MD, López-Linares K, Posada J. Blood flow visualization from 4D flow Magnetic Resonance Imaging using the ISO X3D standard. In: Proceedings of the 29th International ACM Conference on 3D Web Technology. New York, NY, USA: Association for Computing Machinery; 2024:1-5.
16. Roberts GS, Hoffman CA, Rivera-Rivera LA, Berman SE, Eisenmenger LB, Wieben O. Automated hemodynamic assessment for cranial 4D flow MRI. Magn Reson Imaging 2023;97:46-55. [Crossref] [PubMed]
17. Schrauben E, Wåhlin A, Ambarki K, Spaak E, Malm J, Wieben O, Eklund A. Fast 4D flow MRI intracranial segmentation and quantification in tortuous arteries. J Magn Reson Imaging 2015;42:1458-64. [Crossref] [PubMed]
18. Corrado PA, Wentland AL, Starekova J, Dhyani A, Goss KN, Wieben O. Fully automated intracardiac 4D flow MRI post-processing using deep learning for biventricular segmentation. Eur Radiol 2022;32:5669-78. [Crossref] [PubMed]
19. Merton R, Bosshardt D, Strijkers GJ, Nederveen AJ, Schrauben EM, van Ooij P. Reproducibility of 3D thoracic aortic displacement from 3D cine balanced SSFP at 3 T without contrast enhancement. Magn Reson Med 2024;91:466-80. [Crossref] [PubMed]
20. Merton R, Bosshardt D, Nederveen A, Strijkers G, Ooij PV, Schrauben E. Kiosk 3R-FC-06 - Improving Robustness of 3D Cine Bssfp at 3 T. J Cardiovasc Magn Reson 2024;26:S1-926.
21. Gottwald LM, Peper ES, Zhang Q, Coolen BF, Strijkers GJ, Nederveen AJ, van Ooij P. Pseudo-spiral sampling and compressed sensing reconstruction provides flexibility of temporal resolution in accelerated aortic 4D flow MRI: A comparison with k-t principal component analysis. NMR Biomed 2020;33:e4255. [Crossref] [PubMed]
22. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021;18:203-11. [Crossref] [PubMed]
23. Schrauben EM. Amsterdam UMC FlowProcessingTool. 1.0 ed. Github 2024. Available online: https://github.com/schrau24/FlowProcessing
24. Schrauben EM, Aalbregt E. Automated abdominal aortic aneurysm (AAA) 4D flow processing pipeline. 1.0 ed. Github 2025. Available online: https://github.com/schrau24/AAA_4Dflow_auto
25. Loecher M, Schrauben E, Johnson KM, Wieben O. Phase unwrapping in 4D MR flow with a 4D single-step laplacian algorithm. J Magn Reson Imaging 2016;43:833-42. [Crossref] [PubMed]
26. Ong F, Uecker M, Tariq U, Hsiao A, Alley MT, Vasanawala SS, Lustig M. Robust 4D flow denoising using divergence-free wavelet transform. Magn Reson Med 2015;73:828-42. [Crossref] [PubMed]
27. Chatterjee D, Shen TC, Mukherjee P, Lee S, Garrett JW, Zacharias N, Pickhardt PJ, Summers RM. Automated detection of incidental abdominal aortic aneurysms on computed tomography. Abdom Radiol (NY) 2024;49:642-50. [Crossref] [PubMed]
28. Barker AJ, Markl M, Bürk J, Lorenz R, Bock J, Bauer S, Schulz-Menger J, von Knobelsdorff-Brenkenhoff F. Bicuspid aortic valve is associated with altered wall shear stress in the ascending aorta. Circ Cardiovasc Imaging 2012;5:457-66. [Crossref] [PubMed]
29. Frydrychowicz A, Stalder AF, Russe MF, Bock J, Bauer S, Harloff A, Berger A, Langer M, Hennig J, Markl M. Three-dimensional analysis of segmental wall shear stress in the aorta by flow-sensitive four-dimensional-MRI. J Magn Reson Imaging 2009;30:77-84. [Crossref] [PubMed]
30. Rinaudo A, Pasta S. Regional variation of wall shear stress in ascending thoracic aortic aneurysms. Proc Inst Mech Eng H 2014;228:627-38. [Crossref] [PubMed]
31. Potters WV, Marquering HA, VanBavel E, Nederveen AJ. Measuring Wall Shear Stress Using Velocity-Encoded MRI. Curr Cardiovasc Imaging Rep 2014;7:9257.
32. van Ooij P, Powell AL, Potters WV, Carr JC, Markl M, Barker AJ. Reproducibility and interobserver variability of systolic blood flow velocity and 3D wall shear stress derived from 4D flow MRI in the healthy aorta. J Magn Reson Imaging 2016;43:236-48. [Crossref] [PubMed]
33. Tanweer O, Wilson TA, Metaxa E, Riina HA, Meng H. A comparative review of the hemodynamics and pathogenesis of cerebral and abdominal aortic aneurysms: lessons to learn from each other. J Cerebrovasc Endovasc Neurosurg 2014;16:335-49. [Crossref] [PubMed]
34. Cañadas V, Vilacosta I, Luaces M, Bustos A, Ferreirós J, Aragoncillo P, Pérez de Isla L, Rodríguez E. Thrombosis of an Apparently Normal Thoracic Aorta and Arterial Embolism. Rev Esp Cardiol 2008;61:196-200.
35. Bustamante M, Viola F, Engvall J, Carlhäll CJ, Ebbers T. Automatic Time-Resolved Cardiovascular Segmentation of 4D Flow MRI Using Deep Learning. J Magn Reson Imaging 2023;57:191-203. [Crossref] [PubMed]
36. Marin-Castrillon DM, Lalande A, Leclerc S, Ambarki K, Morgant MC, Cochet A, Lin S, Bouchot O, Boucher A, Presles B. 4D segmentation of the thoracic aorta from 4D flow MRI using deep learning. Magn Reson Imaging 2023;99:20-5. [Crossref] [PubMed]
37. Merton R, Bosshardt D, Strijkers GJ, Nederveen AJ, Schrauben EM, van Ooij P. Assessing aortic motion with automated 3D cine balanced steady state free precession cardiovascular magnetic resonance segmentation. J Cardiovasc Magn Reson 2024;26:101089. [Crossref] [PubMed]
38. Hurd ER, Iffrig E, Jiang D, Oshinski JN, Timmins LH. Flow-based method demonstrates improved accuracy for calculating wall shear stress in arterial flows from 4D flow MRI data. J Biomech 2023;146:111413. [Crossref] [PubMed]
39. Casciaro ME, Pascaner AF, Guilenea FN, Alcibar J, Gencer U, Soulat G, Mousseaux E, Craiem D. 4D flow MRI: impact of region of interest size, angulation and spatial resolution on aortic flow assessment. Physiol Meas 2021;
40. Aalbregt E, Rijken L, Nederveen A, van Ooij P, Yeung KK, Jongkind V. Quantitative Magnetic Resonance Imaging to Assess Progression and Rupture Risk of Aortic Aneurysms: A Scoping Review. J Endovasc Ther 2025;32:929-45. [Crossref] [PubMed]