Local pressure as a dominant hemodynamic driver of wall enhancement in anterior communicating artery aneurysms: a facet-level computational fluid dynamics and vessel wall imaging analysis

Introduction

Unruptured intracranial aneurysms (UIAs) represent a significant public health burden, with an estimated global prevalence of approximately 3% in the adult population (1). While the majority of UIAs remain quiescent, aneurysmal subarachnoid hemorrhage (aSAH) entails devastating consequences: despite advances in microsurgical and endovascular management, aSAH continues to carry high rates of mortality and permanent disability, driven by early brain injury, rebleeding, and delayed cerebral ischemia (2).

Given that prophylactic intervention itself carries a non-negligible risk of morbidity, often exceeding 5% even in modern cohorts (3), clinical decision-making hinges on accurately identifying the minority of lesions destined to rupture. Current risk stratification relies heavily on scoring systems such as the PHASES score, which prioritizes aneurysm size and population history. However, the reliability of these morphological-based metrics has been increasingly challenged. A recent retrospective analysis by Krystkiewicz et al. revealed that the PHASES score failed to distinguish between ruptured and unruptured aneurysms in a real-world cohort, with nearly identical risk scores observed in both groups (4). Crucially, they reported that over 50% of ruptured aneurysms were small (<6 mm) and would have been classified as low-risk by current guidelines. This paradox highlights a critical diagnostic gap: traditional size-based metrics systematically underestimate the lethal potential of small aneurysms, underscoring the urgent need for physiological biomarkers that reflect vessel wall instability beyond simple geometry.

Aneurysm wall enhancement (AWE) on high-resolution vessel wall imaging (VWI) has emerged as a promising surrogate marker for wall instability. Recent longitudinal meta-analyses confirm that AWE serves as a robust predictor of aneurysm progression, with enhancing lesions carrying a 3.6-fold higher risk of instability compared to non-enhancing ones (5). Histologically, this radiological phenomenon has been substantiated by a recent systematic review correlating VWI with surgical specimens, which linked enhancement to macrophage infiltration (CD68/MPO) and neovascularization (CD34) within the vessel wall (6). Beyond the aneurysm itself, AWE has also been linked to the burden of coexisting intracranial atherosclerotic plaque on high-resolution VWI, with higher plaque grades associated with more pronounced wall enhancement, further supporting AWE as a marker of active, inflammation-related wall pathology (7). However, the underlying pathophysiological mechanisms driving this enhancement remain poorly understood. Specifically, it is unclear whether AWE represents a uniform inflammatory response or a localized pathology driven by specific mechanical forces.

Anterior communicating artery (ACoA) aneurysms represent a unique and physiologically volatile model to investigate this mechanobiological interaction. Anatomically, the ACoA complex is subject to the “firehose nozzle effect” driven by upstream vessel tapering. As demonstrated by Lauric et al., A1 segments leading to aneurysms exhibit significant caliber narrowing, which accelerates inflow and subjects the bifurcation to distinctively high focal pressure and velocity (8).

At the bifurcation apex itself, Guo et al. utilized virtual aneurysm removal techniques to reveal a paradoxical hemodynamic environment: while the flow impingement zone bears the highest total pressure, it simultaneously creates a focal stagnation point characterized by critically low wall shear stress (WSS) and high gradients (9). Consistently, using the same virtual aneurysm removal approach at the internal carotid artery bifurcation, Li et al. found that hemodynamic stresses on the aneurysm dome were markedly lower than those on the reconstructed bifurcation apex, supporting the view that aneurysm formation itself serves to relieve the abnormally elevated stresses generated by direct flow impingement (10). Furthermore, recent computational studies by Tian et al. highlight that ACoA hemodynamics are profoundly influenced by the global configuration of the Circle of Willis (e.g., cross-flow from A1 dominance), necessitating high-fidelity global modeling to accurately map these flow disturbances (11). However, while these studies have characterized the flow field, the link between these complex hemodynamic vectors and the biological response of the vessel wall (AWE) remains undefined. Specifically, it is unclear whether these stagnation zones are merely physical phenomena or the direct triggers for the inflammatory remodeling observed on magnetic resonance imaging (MRI).

To overcome the limitations of global averaging, pioneering studies utilizing ultra-high-field 7T MRI have introduced point-wise mapping techniques. Notably, Hadad et al. demonstrated that localized WSS is negatively correlated with wall enhancement intensity at the voxel level, a relationship often obscured by whole-aneurysm analysis (12). However, these insights have largely been restricted to research-grade 7T imaging, limiting their generalizability to standard clinical workflows (3T MRI). Furthermore, few studies have applied this high-fidelity mapping specifically to ACoA aneurysms, in which complex anatomical variation challenges standard registration algorithms. Moreover, while prior studies have predominantly focused on signal intensity, the precise spatial coupling between hemodynamics and quantitative structural wall remodeling, a potential marker of neovascular burden, remains to be fully elucidated in this complex vascular territory.

Therefore, this study developed a robust sub-voxel co-registration pipeline to spatially align computational fluid dynamics (CFD) simulations with clinical 3T-VWI data. By integrating this rigorous mapping with machine learning interpretation, we aimed to comprehensively characterize the local hemodynamic forces driving both the initiation and severity of wall thickening. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0249/rc).

Methods

Study design and patient population

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Zhujiang Hospital, Southern Medical University (IRB number: 2023-KY-038-03) and individual consent for this retrospective analysis was waived. The study is registered at ClinicalTrials.gov (NCT06447714). Clinical and imaging data were retrospectively collected from a single academic medical center between January 2023 and June 2024.

The distinctive hemodynamics of ACoA aneurysms, particularly the complex flow patterns arising from the confluence of bilateral anterior cerebral arteries, provide an ideal setting to investigate the relationship between vessel wall features and hemodynamic stimuli.

Patients were eligible for inclusion if they met the following criteria: (I) age ≥18 years; (II) diagnosis of saccular IA at the ACoA confirmed by DSA or CTA; (III) availability of both three-dimensional (3D) time-of-flight MRA (TOF-MRA) and high-resolution VWI including pre- and post-contrast T1-weighted sequences; (IV) adequate image quality without significant motion artifacts, suitable for three-dimensional reconstruction; and (V) complete clinical data with documented risk factor profiles.

Exclusion criteria comprised: (I) dissecting or fusiform aneurysms; (II) aneurysms with diameter <2 mm or >10 mm; (III) previous surgical or endovascular treatment; (IV) concurrent major cerebrovascular anomalies; and (V) inadequate image quality for segmentation.

Aneurysms smaller than 2 mm were excluded because, at the reconstructed voxel size of 0.60 mm × 0.60 mm × 0.60 mm, they span fewer than approximately three voxels across their dominant dimension and therefore cannot support reliable lumen-adventitia segmentation or the sub-voxel sampling step of the Vector-Integrated Surface Parametrization (VISP) pipeline. This threshold reflects an imaging-resolution constraint rather than a clinical severity criterion. The upper bound of 10 mm reflects the prespecified focus of this study on small-to-medium ACoA lesions, which constitute the clinically most uncertain subgroup under current size-based risk stratification.

MR image acquisition

TOF-MRA images were acquired using the following parameters: repetition time/echo time (TR/TE) =20/3.5 ms, field of view (FOV) =200 mm × 180 mm, voxel size =0.50 mm × 0.75 mm, flip angle =18°, and a total acquisition time of 2 min 55 s. Aneurysm wall imaging was then performed using a 3D black-blood T1-weighted volumetric turbo spin echo (T1-VISTA) sequence. The acquisition parameters were: TR/TE =800/20 ms, FOV =250 mm × 161 mm, acquisition matrix =416×269, reconstructed voxel size =0.60 mm × 0.60 mm × 0.60 mm, 80 contiguous slices, flip angle =90°, number of signal average =1, resulting in a total scan time of 6 min 2 s. Finally, post-contrast VWI was performed 2 minutes after intravenous injection of Gd-DTPA (Magnevist; Bayer Schering Pharma, Berlin, Germany) at a dose of 0.1 mmol/kg.

Segmentation protocol and spatial co-registration

The MRI series were segmented and reconstructed into 3D models using Mimics Research (V21.0, Materialise NV, Leuven, Belgium). Two neuroradiologists (H.C. and F.X., with 5 and 10 years of experience, respectively) independently participated in the 3D model generation with a high resolution of surface mesh (mean edge length: 0.16±0.05 mm).

Two source image sets were used for reconstruction: the 3D TOF-MRA (Figure 1, A1) and the post-contrast T1-VWI (Figure 1, A2). From the TOF-MRA, the cerebral vascular tree model (Model-1) was extracted to define the hemodynamic flow domain (Figure 1, A3). From the post-contrast T1-VWI, two distinct structures were then segmented (Figure 1, A4): (I) an anatomical reference model (white, Model-2), encompassing the aneurysm sac and parent arteries; and (II) the peri-aneurysmal hyperintense region (blue, Model-3), segmented as a volume of interest to capture the full extent of potential wall enhancement.

Figure 1 Workflow for segmentation, spatial co-registration, and enhancement quantification of intracranial aneurysm wall imaging. (A) Model reconstruction and registration. (A1) Axial 3D TOF-MRA demonstrating the IA. (A2) Corresponding post-contrast T1-VWI slice. (A3) Three-dimensional vascular tree model reconstructed from TOF-MRA. (A4) Segmented models from T1-VWI: anatomical reference model comprising the IA sac and parent arteries (white) and the peri-aneurysmal hyperintense volume of interest (blue). (A5) Spatially registered vascular tree following rigid transformation to the T1-VWI coordinate space. (A6) Final composite model after Boolean intersection, depicting the registered vasculature (red) with the surface-confined hyperintense region (blue). (B) Mapping from IA surface to primary enhancement. (B1) Schematic of the ray-casting methodology, with rays projected from the aneurysm centroid through each surface vertex. (B2) Surface mapping of primary enhancement distribution, with color scale representing normalized signal intensity. (B3) Cross-sectional visualization of the vessel wall and corresponding enhancement region. (C) Enhancement quantification. (C1) Illustration of GWT and EWT measurements. GWT was defined as the Euclidean distance between the inner (P-inner) and outer (P-outer) wall boundaries. EWT was calculated as: EWT = GWT × (N-enhanced/N-total), where N-enhanced and N-total represent the number of enhanced and total sampled voxels, respectively. (C2) Representative visualization of the quantified enhancement region (purple) overlaid on the composite vascular model. Three reference positions (center, inner, and outer) were computationally defined along each ray path. Given the negligible separation between center and inner wall positions (on the order of 10⁻³ mm), these two points are represented as a single combined point in panels B1 and C1 for visualization purposes. 3D, three-dimensional; CR, contrast ratio; EWT, enhanced wall thickness; GWT, geometric wall thickness; IA, intracranial aneurysm; N-enhanced, number of enhanced voxels; N-total, number of total sampled voxels; P-inner, inner wall boundary; P-outer, outer wall boundary; T1-VWI, T1-weighted vessel wall imaging; TOF-MRA, time-of-flight magnetic resonance angiography.

To establish spatial correspondence, we performed rigid registration to spatially align the TOF-MRA vascular tree with the T1-VWI anatomical reference model (Figure 1, A5). Registration accuracy was confirmed by 3D deviation analysis (Geomagic Control X 2021, 3D Systems), achieving sub-millimeter alignment precision as detailed in Table S1. Intersection of the registered model with the peri-aneurysmal hyperintense region via a Boolean operation yielded the aneurysm-surface-confined high-signal region. This extracted region was then combined with the registered vasculature to generate a composite model (Figure 1, A6), which was employed for quantitative analysis.

To validate the segmentation protocol, a blinded reliability analysis was conducted on a randomly selected subset (40% of the cohort). Both radiologists independently segmented the required models, with volumetric agreement quantified using the Dice similarity coefficient. Initial discrepancies were resolved by consensus review before proceeding with collaborative segmentation for the remaining cohort.

Quantification of AWE

We developed a standardized three-step pipeline to quantify AWE by integrating signal intensity and wall thickness metrics.

To normalize signal intensity for inter-patient heterogeneity, the post-contrast T1-VWI signal of the pituitary stalk served as the internal reference (13,14). The contrast ratio (CR) was defined as SI_voxel/mean (SI_stalk). To define enhancement, a CR threshold of ≥0.6 was established following a sensitivity analysis that balanced detection sensitivity and specificity (Tables S2,S3). The CR ≥0.6 threshold was selected as the primary operating point on the basis of a formal sensitivity analysis across CR =0.5, 0.6, and 0.7 (Table S4); CR =0.5 yielded a non-informative class prevalence (≥84% enhanced), whereas CR =0.7 preserved the direction of the primary finding with a doubled effect magnitude (see Results and Supplementary). Enhancement metrics were computed at the facet level. For each triangular facet on the aneurysm sac model, a ray was cast from its centroid along the outward surface normal toward the peri-aneurysmal hyperintense region (Figure 1, B1). Figure 1, B2,B3, depict the distribution of the distance matrix across the aneurysm sac in both full-view and cross-sectional profiles. Inner and outer boundary intersections were identified using the Möller-Trumbore ray-triangle algorithm, and the Euclidean distance between these intersection points defined the local geometric wall thickness (GWT).

To ensure robust quantification of enhancement across varying wall thicknesses, we implemented an adaptive high-density sampling strategy along each ray segment. Sampling density was dynamically optimized to the local GWT instead of using a fixed interval, guaranteeing that resolution always exceeded the native MRI voxel size. This minimized partial volume effects and captured fine intra-wall intensity variations. Signal intensities at these sampled points (N) were then trilinearly interpolated from the co-registered T1-VWI dataset. The complete two-stage algorithmic specification, encompassing the ray-casting computation of GWT (Stage A) and the sub-voxel intensity sampling for enhanced wall thickness (EWT) (Stage B), is provided in Appendix 1.

Two facet-level AWE metrics were derived as illustrated in Figure 1, C1: (I) EWT, defined as the effective enhancement thickness calculated as:

EWT=GWT×(NenhancedNtotal)[1]

where N_enhanced is the number of sample points with CR ≥0.6, N_total is the total number of samples along the ray; and (II) Mean CR, representing the average intensity of the enhanced portion, calculated as:

MeanCR=∑CRenhancedNenhanced[2]

where ∑CRenhanced represents the cumulative CR of all enhanced sample points. If a ray failed to intersect both boundaries or if no samples met the enhancement criterion (N_enhanced=0), both EWT and Mean CR were assigned a value of zero. The final volumetric distribution of these enhanced sampling points was visualized on the composite model, with the specific sampling density highlighted in the magnified view of Figure 1, C2.

Hemodynamic simulation and multi-modal data fusion

High-quality volume meshes (approximately 1.7 to 4.1 million tetrahedral elements) were generated from the TOF-derived models in ANSYS ICEM CFD. The Circle of Willis geometry was preserved to the greatest extent visible on TOF-MRA to avoid artificially prescribing flow directions at the communicating arteries. A boundary layer of six prism layers with a growth ratio of 1.2 was created to resolve near-wall velocity gradients.

Transient simulations were run in ANSYS CFX (2020 R2), modeling blood as a laminar, incompressible, non-Newtonian fluid (ρ=1,060 kg/m³). The Carreau-Yasuda model was reported as the most suitable model in simulating the blood properties of cerebral arteries, where the detailed parameter settings and the constitutive model were delineated in our previous work (15,16). Pulsatile inlet flow and outlet pressure waveforms were derived from the @neufuse software, which encapsulates a one-dimensional model of the human vascular tree comprising 63 major arteries calibrated to a healthy-adult population (17). The waveform database provides a representative cardiac cycle of 0.80 s (corresponding to a heart rate of 75 bpm) with a combined cerebral inflow of approximately 500 mL/min at the proximal arterial trunks. Because these waveforms are population-averaged rather than patient-specific, the absolute magnitudes of the computed pressure and WSS fields carry an estimated inter-individual uncertainty on the order of 20 to 30 percent, consistent with literature-reported coefficients of variation of 15 to 25 percent for flow velocity and 10 to 15 percent for peak velocity. Within each patient, the Circle of Willis geometry was preserved to the greatest extent visible on TOF-MRA, so that the flow direction and magnitude through the communicating arteries emerged from the simulation rather than being prescribed. Simulations used a timestep of 0.008 s over three cardiac cycles; results from the final cycle were analyzed after periodic steady state.

To integrate multi-modal data, a custom integrated methodology termed VISP was implemented. All variables were mapped onto the centroid of each triangular facet to establish a unified analytical framework. The two AWE metrics (EWT and MeanCR) were assigned directly to the facet centroid. Hemodynamic parameters were linearly interpolated from the three mesh vertices to the facet centroid using barycentric coordinates.

Statistical analysis

All statistical analyses were conducted in Python (v3.9) using statsmodels and scikit-learn libraries, with a fixed random seed [42] for reproducibility. Continuous hemodynamic variables were Z-score standardized prior to regression and machine learning modeling. For the patient presenting with two ACoA aneurysms, surface segments from both lesions were assigned to a single PatientID to ensure consistent handling of within-patient clustering across all analyses; the final cohort therefore comprises 24 unique patients harboring 25 ACoA aneurysms.

To identify hemodynamic drivers of wall enhancement while explicitly accounting for the hierarchical structure of the data (segments nested within aneurysms within patients), four complementary analytical frameworks were applied in parallel: (I) intra-patient paired bootstrap tests; (II) a multivariate linear mixed model (LMM) with patient-level random intercepts; (III) generalized estimating equations (GEEs) with cluster-robust variance; and (IV) ensemble machine-learning models interpreted via Shapley Additive exPlanations (SHAP) values across four cross-validation (CV) regimes.

To control for systemic inter-individual variability, hemodynamic differences between enhanced and non-enhanced segments were assessed using a strictly paired design within each patient (framework I). Intra-patient comparisons were performed using paired bootstrap tests (2,000 resamples) within each of the 13 AWE-positive patients, with enhanced and non-enhanced segments paired within the same aneurysm. P-values were adjusted via the false discovery rate (FDR) method across the 13 hemodynamic parameters tested in parallel. For the patient presenting with multiple aneurysms, only the dominant lesion was included in this paired analysis to prevent pseudoreplication.

To identify independent determinants of wall thickening magnitude among enhanced segments (framework II; 12,473 enhanced segments in total, reduced to a cluster-balanced subset of 2,573 after capping each patient at 300 segments), a multivariate LMM was fitted with patient-level random intercepts to explicitly decompose residual variance into between-patient and within-patient components. The fixed-effect specification included an explicit intercept term and the Z-scored hemodynamic features. The analysis proceeded in two stages. First, the association of each hemodynamic parameter with EWT was assessed using univariate LMMs; only parameters reaching FDR-corrected significance were advanced to the multivariate candidate pool. Multicollinearity among candidates was then controlled by excluding variables with variance inflation factor (VIF) >5. The final optimized model was selected via forward stepwise inclusion based on the Akaike information criterion (AIC), and parameter stability was validated by 1,000 non-parametric bootstrap resamples. To verify that fixed-effect estimates were not driven by a few patients contributing disproportionately many segments, a cap-sensitivity analysis was performed across four per-patient sampling caps (300, 500, 1,000, and 10,000 segments, with the largest cap effectively uncapped given the per-patient segment counts in the AWE-positive subset), reporting the direction, magnitude, and statistical significance of each retained feature at every cap to evaluate robustness against cluster-balancing choices (Table S5,S6).

For the binary AWE-presence outcome across the full segment cohort (framework III; n=157,284), the primary inferential analysis was a GEE model with a logit link, binomial family, and exchangeable working correlation clustered by PatientID. GEE provides a population-averaged estimator that natively accounts for within-patient hierarchical dependence via cluster-robust standard errors (SEs) and does not require regularization for identifiability. Cross-patient generalization of AWE presence was disclosed separately under strict patient-level Stratified GroupKFold CV. The candidate feature pool followed the same dual-criterion screening logic used for the multivariate LMM (univariate significance and VIF <5). For completeness, the original L2-regularized multivariate logistic model was retained as a supplementary collinearity-diagnostic comparison, reported alongside VIF and partial-regression analyses to explain the apparent coefficient sign reversals observed under L2-regularized logistic regression (Tables S7-S9 and Figures S1,S2).

To capture non-linear hemodynamic interactions and quantitatively rank feature importance, ensemble machine-learning models (Random Forest and XGBoost) were trained on the enhanced-segment cohort (framework IV). The feature space comprised the hemodynamic parameters reaching univariate LMM significance. Four complementary CV strategies were applied to address scientifically distinct questions: (I) segment-level five-fold CV, retained as a co-primary analysis of within-lesion mechanobiological coupling rather than as a claim of cross-patient generalization; (II) patient-centered residualized analysis, in which each patient’s mean EWT was subtracted prior to modeling to isolate the within-patient hemodynamic-to-EWT gradient; and (III) patient-level GroupKFold CV, in which entire patients were held out per fold as a strict out-of-sample estimate of cross-patient generalizability. Hyperparameters were optimized via nested grid search within each CV regime, and global feature importance was quantified using SHAP; ranking stability across CV regimes was reported as a supplementary robustness analysis (Tables S10,S11, Figure S3); (IV) patient-centered GroupKFold with median centering as a sensitivity check on the centering statistic

Results

Patient characteristics and technical validation

Out of an initial cohort of 123 patients screened for intracranial aneurysms (IAs), 99 were excluded due to non-ACoA location (n=68), insufficient image quality for sub-voxel segmentation (n=20), or other exclusion criteria (n=11), yielding a final homogeneous cohort of 24 patients harboring 25 unruptured ACoA aneurysms (Figure 2). Among these 25 aneurysms, focal AWE was identified in 14 aneurysms (56.0%) corresponding to 13 patients (the AWE-positive group), while 11 aneurysms showed no enhancement (the AWE-negative group). One patient harbored two aneurysms, both of which exhibited wall enhancement (Table 1)

Figure 2 Patient selection flow diagram in accordance with TRIPOD+AI reporting requirements. Of 123 patients initially screened for intracranial aneurysms, 99 were excluded due to non-ACoA aneurysm location (n=68), insufficient image quality for sub-voxel segmentation (n=20), or other criteria (n=11) comprising dissecting/fusiform morphology, aneurysm diameter <2 mm or >10 mm, prior surgical or endovascular treatment, and concurrent major cerebrovascular anomalies. The final cohort comprised 24 patients harboring 25 ACoA aneurysms (one patient harbored two aneurysms, both AWE-positive), stratified into AWE-positive (13 patients, 14 aneurysms; primary analysis cohort) and AWE-negative (11 patients, 11 aneurysms; control cohort) subgroups based on a contrast ratio threshold of CR ≥0.6. Four hierarchical analytical frameworks were applied to address within-patient clustering: (I) intra-patient paired bootstrap testing within each aneurysm; (II) linear mixed model with patient-level random intercepts for EWT severity; (III) generalized estimating equations with cluster-robust variance for AWE presence; and (IV) random forest with SHAP-based interpretation across four cross-validation regimes (segment-level, patient-level Stratified GroupKFold, patient-centered mean, and patient-centered median). ACoA, anterior communicating artery; AWE, aneurysmal wall enhancement; CR, contrast ratio; EWT, enhanced wall thickness; SHAP, SHapley Additive exPlanations.

Baseline demographic characteristics, including age (P=0.600) and sex (P=0.217), were comparable between the AWE-positive and AWE-negative groups. Regarding clinical risk factors, no significant differences were observed in the prevalence of hypertension, diabetes, dyslipidemia, smoking status, or history of stroke (P>0.05). Notably, alcohol consumption was statistically more frequent in the AWE-negative group (P=0.031).

Morphological analysis revealed significant differences between the two groups. Aneurysms in the AWE-positive group were significantly larger in maximal diameter (D_max: 6.1±2.2 vs. 3.8±2.2 mm, P=0.013). Furthermore, both Aspect Ratio (0.9±0.3 vs. 0.6±0.2, P=0.011) and Size Ratio (2.5±1.0 vs. 1.6±0.9, P=0.015) were significantly higher in the AWE-positive group, indicating that aneurysms with wall enhancement tended to have more irregular, elongated morphologies.

Anatomical variation of the ACoA complex was also significantly associated with wall enhancement (P=0.046). Specifically, all aneurysms associated with Unilateral A1 dominance (n=4) exhibited AWE (100%). In contrast, AWE was present in 63.6% of cases with Bilateral A1 co-dominance and only 30.0% of cases with Normal anatomy.

The VISP pipeline demonstrated excellent reproducibility. Inter-observer reliability for Model-1, Model-2 and Model-3 reconstructions yielded the Dice similarity coefficient of 0.94±0.03, 0.92±0.04, and 0.88±0.06, respectively. Three-dimensional deviation analysis confirmed sub-millimeter registration accuracy between the Model-1 and Model-2, with a mean alignment error of ≤0.3 mm and root-mean-square deviation of approximately 0.25 mm (Table S1).

Intra-patient hemodynamic signature of IA wall enhancement

To isolate regional hemodynamic determinants of AWE while controlling for systemic confounders, we performed paired comparisons between enhanced (EWT >0) and non-enhanced (EWT =0) wall segments within the 13 patients included in the final paired cohort. All hemodynamic P-values reported below are FDR-adjusted across the 13 parameters; raw P-values are reported with explicit notation (P) for context only. A comprehensive panel of 13 hemodynamic parameters was analyzed, organized into three distinct temporal categories:

• Peak systolic phase metrics: Pressure_peak, WSS_peak, wall shear stress gradient (WSSG_peak), aneurysm formation index (AFI_peak), and Viscosity_peak;
• Mid-systolic deceleration (MSD) phase metrics: AFI_MSD, WSS_MSD, Viscosity_MSD, and Vorticity_MSD;
• Time-averaged (cycle-wide) metrics: time-averaged WSS (TAWSS), oscillatory shear index (OSI), gradient oscillation number (GON), and time-averaged viscosity (TA-Viscosity).

Qualitative and quantitative analyses revealed that enhanced wall segments are characterized by a distinct hemodynamic microenvironment dominated by flow stagnation and reduced mechanical forces. As visualized in the representative 3D hemodynamic maps, the enhanced wall region (upper row, Figure 3A-3C) and the non-enhanced wall region (lower row, Figure 3D-3F) are shown for Pressure_peak, TAWSS, and WSS_peak, respectively; the region of wall enhancement (delineated by black contour lines) shows a striking spatial concordance with zones of low hemodynamic energy.

Figure 3 Representative 3D hemodynamic maps illustrating the spatial concordance between aneurysm wall enhancement and low-energy hemodynamic zones. The upper row (A-C) specifically displays the Enhanced Wall Region (delineated by black contour lines), which spatially corresponds to sheltered, low-energy zones. The lower row (D-F) displays the Global Aneurysm Sac, serving as the reference for the surrounding Non-Enhanced Regions. (A,D) Peak systolic pressure: the enhancement region (A) is strictly confined to the low-pressure zone, contrasting with the high-pressure impingement zone observed in the global reference (D). (B,E) TAWSS: enhancement aligns perfectly with regions of flow stagnation (low TAWSS), while the surrounding non-enhanced wall typically experiences higher shear. (C,F) Peak systolic WSS: the enhanced wall experiences significantly lower instantaneous shear stress compared to the global average. 3D, three-dimensional; TAWSS, time-averaged wall shear stress; WSS, wall shear stress.

Specifically, at the peak systolic phase, enhanced regions exhibited significantly lower local Pressure_peak (mean difference =−35.25 Pa; P_FDR =0.02) compared with non-enhanced regions. Visual inspection confirms that AWE predominantly localizes to geometrically sheltered low-pressure zones in the enhanced region, in contrast to the high-pressure impingement zones of the non-enhanced region (Figure 3A,3D). Similarly, WSS_peak was significantly reduced in enhanced segments (mean difference =−3.97 Pa; P_FDR <0.001), with 3D mapping showing the enhancement strictly confined to low shear stress regions (Figure 3C,3F).

Regarding cycle-wide metrics, the hemodynamic profile of flow stagnation persisted throughout the cardiac cycle. Enhanced regions exhibited a consistent reduction in TAWSS (mean difference =−0.79 Pa). While strictly marginal after FDR correction (P_FDR =0.061), the unadjusted analysis revealed a significant reduction (P=0.014) that directionally aligns with the profound WSS_peak deficit observed at peak systole.

Similarly, secondary rheological and oscillatory disturbances were noted in the raw analysis. Specifically, enhanced regions showed a trend toward elevated TA-Viscosity (P=0.021) and OSI (P=0.032). However, unlike the peak systolic metrics, these cycle-averaged parameters did not retain statistical significance after correcting for multiple comparisons (Table 2).

A formal CR threshold sensitivity analysis across CR =0.5, 0.6, and 0.7 confirmed the robustness of this primary hemodynamic signature (Table S4). At CR =0.5, the contrast between enhanced and non-enhanced classes became non-informative due to ≥84% segment-level prevalence (ΔPressure_peak =+1.99 Pa; P_FDR =0.996); at CR =0.6 (primary), ΔPressure_peak =−35.25 Pa (P_FDR =0.02); at CR =0.7, the effect magnitude approximately doubled (ΔPressure_peak =−61.99 Pa; P_FDR <0.001) with preserved direction, demonstrating a monotonic dose-response that supports CR =0.6 as a conservative primary threshold.

Collectively, these findings define the primary hemodynamic signature of AWE: a significant, spatially-concordant reduction in both instantaneous (peak systolic) and sustained (time-averaged) mechanical loading, identifying the enhanced wall as a region of hemodynamic stagnation.

Segment-level correlations between hemodynamic parameters and wall enhancement metrics

To characterize the spatial correspondence between local hemodynamics and the severity of wall pathology, Pearson correlation coefficients were computed specifically across the 12,473 segments that exhibited enhancement (EWT >0). This restricted analysis focuses on determining which hemodynamic factors drive the extent and intensity of pathology within established lesion zones, rather than differentiating between healthy and diseased walls. Detailed statistical results are provided in Tables S12,S13.

EWT, reflecting the anatomical extent of wall thickening within the lesion, exhibited the highest positive correlation coefficient with TA-Viscosity (r=0.217, P_FDR <0.001). While the effect size indicates a modest linear relationship, viscosity emerged as the most significant hemodynamic correlate of wall thickening among all parameters evaluated (Figure 4A). Additional positive associations were observed with GON (r=0.125) and OSI (r=0.121), suggesting that disturbed flow patterns contribute to the extent of thickening. Conversely, shear-related parameters exhibited consistent, albeit weak, negative correlations, including AFI_peak (r=−0.151) and WSS_MSD (r=−0.100). Pressure_peak showed a negligible negative correlation with EWT (r=−0.025, P_FDR =0.004).

Figure 4 Statistical identification of hemodynamic drivers governing the severity of aneurysm wall enhancement. Analyses were restricted to the subset of surface segments exhibiting focal enhancement (n=12,473) to identify hemodynamic factors modulating lesion severity. (A,B) Univariate correlation hierarchy. Pearson correlation coefficients (r) ranking hemodynamic parameters by their linear association with wall pathology. Bars are color-coded by direction (teal: positive; maroon: negative). (A) EWT. Time-Averaged Viscosity exhibits the strongest positive correlation (r=0.217), suggesting flow stagnation contributes to anatomical thickening. Shear-related metrics (e.g., AFI, WSS_MSD) show consistent negative associations. (B) MeanCR. Local pressure emerges as the primary negative correlate (r=−0.230), indicating that local hypotension is more sensitively linked to inflammatory intensity than to wall thickness. (C) Independent hemodynamic predictors of EWT severity. Standardized β coefficients from the multivariate LMM fitted on a cluster-balanced subset of 2,573 segments drawn from the 12,473 enhanced segments (cap =300 segments per patient). Error bars denote analytical 95% CI. Following variance inflation factor screening (threshold >5) and AIC-based stepwise selection, five hemodynamic features were retained, with local Pressure_peak persisting as the dominant independent negative predictor of wall thickening (β=−0.181; 95% CI: −0.234 to −0.128; P=2.31×10⁻¹¹; 100% bootstrap stability across 1,000 patient-cluster resamples). Color coding denotes patient-cluster bootstrap stability: teal (high, ≥95%), maroon (moderate, 80–95%), and gray (low, <80%). AFI, anisotropy fractional index; AIC, Akaike information criterion; CI, confidence interval; EWT, enhanced wall thickness; GON, gradient oscillatory number; LMM, linear mixed model; MeanCR, mean contrast ratio; MSD, multi-scale distribution; OSI, oscillatory shear index; SD, standard deviation; TA-Viscosity, time-averaged viscosity; TAWSS, time-averaged wall shear stress; WSS, wall shear stress.

MeanCR, quantifying the inflammatory intensity of the enhanced segments, displayed a distinct correlation hierarchy (Figure 4B). Notably, Pressure_peak emerged as the primary negative correlate of MeanCR (r=−0.230, P_FDR <0.001). Although the absolute correlation strength remains modest, this coefficient is nearly an order of magnitude larger than the correlation between Pressure_peak and EWT, indicating that local hypotension is more sensitively linked to inflammatory intensity than to anatomical thickness. TA-Viscosity maintained a consistent positive association (r=0.244, P_FDR <0.001). These findings highlight that within the complex, multifactorial environment of the aneurysm wall, regions of lower pressure and higher viscosity are statistically associated with more intense inflammatory enhancement.

Independent hemodynamic predictors of enhancement severity: LMMs

To identify the local hemodynamic drivers modulating the severity of wall thickening within established pathological zones, the LMM analysis was restricted to the 12,473 aneurysm surface segments that exhibited focal enhancement (EWT >0) across the 13 AWE-positive patients; this restriction eliminates the confounding influence of non-enhanced vessel walls and focuses the analysis strictly on the determinants of lesion progression magnitude. To prevent disproportionate weight from large-aneurysm clusters and to enforce approximately balanced per-patient contributions, the primary LMM was fitted to a cluster-balanced subset obtained by capping per-patient segment counts at 300 (n=2,573 segments). All models incorporated patient-level random intercepts to account for within-patient hierarchical clustering, with the sensitivity of the primary effect estimates to this sampling choice formally evaluated and reported in the final paragraph.

Univariate LMM screening across the 13 candidate hemodynamic parameters identified nine variables as statistically significant after Benjamini-Hochberg FDR correction (Table S14). Pressure_peak emerged as the parameter with the most substantial standardized association with EWT magnitude [β=−0.155; SE =0.024; 95% confidence interval (CI): −0.201 to −0.108; P_FDR <0.001], confirming that even within the lesional subset deeper local hypotension correlates with greater wall thickness. This univariate result was carried forward into the multivariate model below.

Following VIF screening and AIC-based stepwise selection, the final multivariate LMM retained five hemodynamic features (Figure 4C; Table 3). Pressure_peak persisted as the dominant independent driver of EWT severity (β=−0.181; SE =0.027; 95% CI: −0.234 to −0.128; P=2.31×10⁻¹¹), and patient-cluster bootstrap validation (1,000 resamples) confirmed high stability, with this coefficient remaining statistically significant in 100% of resamples. Four secondary features were retained: AFI_MSD (β=+0.021; SE =0.004; P=1.78×10⁻⁷), WSS_peak (β=+0.025; SE =0.005; P=4.17×10⁻⁶), viscosity (β=−0.020; SE =0.006; P=1.53×10⁻³), and WSSG_peak (β=−0.016; SE =0.005; P=1.66×10⁻³). The intra-class correlation coefficient (ICC) was 0.669 (bootstrap 95% CI: 0.634–0.808), indicating that a substantial fraction of segment-level variance is attributable to between-patient differences and is properly absorbed by the random-intercept structure rather than confounding the fixed-effect estimation.

Cap-sensitivity analysis (Table S5) confirmed that Pressure_peak retained the same direction and statistical significance (all P<10⁻¹⁰) at every sampling cap evaluated [300, 500, 1,000, and 10,000 (effectively uncapped)], spanning a β range of −0.181 to −0.240. Among the secondary features, AFI_MSD, viscosity, and WSSG_peak retained their primary directions and statistical significance across all caps; WSS_peak, however, exhibited a direction reversal at the uncapped level (caP =10,000: β=−0.002; P=0.61), indicating that its apparent positive association at the primary cap is most plausibly a collinearity-driven artefact rather than a robust biological effect. The dominant negative association of Pressure_peak with EWT magnitude is therefore robust to cluster-balancing choice, whereas the directionally minor secondary features are best interpreted as covariates absorbing residual variance under partial confounding rather than as independent biological drivers.

GEE for the presence of wall enhancement

To evaluate the capacity of local hemodynamics to distinguish between healthy and pathological vessel-wall phenotypes while properly accounting for within-patient hierarchical clustering, a GEE model with a logit link, binomial family, and exchangeable working correlation structure clustered by PatientID was fitted to the full segment-level dataset (n=157,284 segments from 24 patients). The GEE model identified Pressure_peak as the dominant independent association with AWE presence (β=−0.5452; robust SE =0.278; robust P=0.0499; 95% CI: −1.090 to −0.0003; odds ratio =0.580, 95% CI: 0.336–0.9997). In contrast to the L2-regularized logistic specification, in which TAWSS and WSS_peak had carried large opposing-sign coefficients (β=+0.607 and β=−1.101, respectively), the corresponding GEE estimates for both parameters were small and non-significant (TAWSS: β=+0.107, robust P=0.31; WSS_peak: β=−0.131, robust P=0.26), consistent with the collinearity-artifact interpretation; full GEE coefficients are reported in Table 4, with diagnostic VIF, partial-regression, and L2-regularized logistic results are reported in Tables S7-S9.

Comparing the GEE analysis (modeling AWE presence on the full cohort) with the multivariate LMM (modeling EWT severity on the enhanced subset), both frameworks converge on Pressure_peak as the dominant negative driver. GEE and LMM handle within-patient clustering by mathematically distinct mechanisms: cluster-robust marginal estimation vs. random-intercept conditional estimation. Their convergent identification of Pressure_peak therefore strengthens the inference that local hypotension is the primary hemodynamic determinant of both the spatial localization (presence) and the magnitude (severity) of wall enhancement in this cohort.

Machine-learning modeling of EWT

To capture non-linear hemodynamic interactions and model the magnitude of EWT, ensemble machine learning models were trained on the 12,473 enhanced segments and evaluated under four distinct CV regimes, each interrogating a different scientific question: segment-level KFold for intra-lesion mechanobiological coupling (co-primary); patient-level Stratified GroupKFold for cross-patient generalization; patient-centered GroupKFold with mean centering (primary) for within-patient hemodynamic-EWT coupling strength; and patient-centered GroupKFold with median centering as a sensitivity check on the centering statistic. Comparative evaluation revealed that the Random Forest regressor achieved superior predictive fidelity compared to XGBoost (Table 5); all results below refer to the optimized Random Forest model. The predictive performance of the Random Forest across all four CV regimes is summarized in Table 6.

Under segment-level 5-fold CV, the model yielded a high coefficient of determination [R²=0.78; mean absolute error (MAE) =0.074 mm; root mean square error (RMSE) =0.103 mm; Figure 6A], indicating that approximately 78% of within-lesion segment-level variance in EWT was explained by local hemodynamic features alone. This metric reflects within-lesion segment-level fit and is retained as a co-primary analysis of within-lesion mechanobiological coupling rather than as a claim of cross-patient generalization. Given the acquired MRI voxel size of isotropic 0.6 mm, the segment-level error margin reflects sub-voxel sampling density achieved through adaptive interpolation rather than imaging resolution beyond the native voxel size.

Figure 6 Random forest mechanobiological modeling of EWT: within-lesion co-primary fit, feature importance, and predictor coupling. (A) Regression fidelity (segment-level co-primary CV). Scatter density plot of actual EWT (x-axis) vs. random forest predicted EWT (y-axis), computed under segment-level 5-fold cross-validation (n=12,473 enhanced segments). The model yielded a high within-lesion coefficient of determination (R²=0.78; MAE =0.074 mm; RMSE =0.103 mm), indicating that local hemodynamic features alone explain approximately 78% of segment-level variance in EWT among enhanced regions. This panel reflects within-lesion mechanobiological coupling under intra-cluster sampling; cross-patient predictive generalization is disclosed separately in Figure 5. (B) Global SHAP feature importance (segment-level co-primary CV). Mean absolute SHAP value ranking eleven hemodynamic predictors by their average contribution to random forest predictions of EWT. Local Pressure_peak emerged as the dominant feature (Rank 1; Mean |SHAP| =0.0301), with the engineered WSS_peak/OSI ratio (Rank 2; Mean |SHAP| =0.0291) closely rivalling its predictive contribution. The Top-3 feature set {Pressure_peak, WSS_peak/OSI, Vorticity_MSD} was preserved across all four cross-validation regimes evaluated (segment-level, patient-level Stratified GroupKFold, and patient-centered mean/median; full sensitivity in Tables S15,S16). (C) Feature correlation matrix. Heatmap of Pearson correlation coefficients among the eleven hemodynamic predictors. The engineered WSS_peak/OSI ratio shows moderate correlation with canonical WSS_peak (r=0.64), and dense coupling is evident among shear-derived metrics. The non-orthogonal feature structure motivates the use of an ensemble random forest model capable of isolating non-linear independent contributions, particularly the dominance of Pressure_peak shown in panels (A) and (B). AFI, anisotropy fractional index; CV, cross-validation; EWT, enhanced wall thickness; MAE, mean absolute error; MSD, multi-scale distribution; OSI, oscillatory shear index; RMSE, root mean square error; SHAP, SHapley Additive exPlanations; WSS, wall shear stress.

To elucidate the biological drivers underlying these predictions, global SHAP feature importance was analyzed under the segment-level co-primary model (Figure 6B; detailed quantitative rankings in Table 7). Consistent with the inferential LMM, Pressure_peak was identified as the dominant feature modulating wall thickening (Rank 1; Mean |SHAP| =0.0301). The engineered WSS_peak to OSI ratio (WSS_peak/OSI) emerged as the second most influential predictor (Rank 2; Mean |SHAP| =0.0291), effectively rivaling Pressure_peak in predictive contribution; this composite metric significantly outperformed the canonical WSS_peak (Rank 5; Mean |SHAP| =0.0112), suggesting that the coupling of shear magnitude and directional instability represents a more potent driver of wall thickening than shear amplitude alone. Mid-systolic flow characteristics, such as Vorticity_MSD (Rank 3), constituted other key determinants. To verify that this mechanistic driver identification is not itself an artefact of the CV strategy, SHAP feature importance was re-computed under the three remaining CV regimes (patient-level Stratified GroupKFold, patient-centered mean, and patient-centered median (full details in Tables S10,S11, and Figure S3). Pressure_peak retained Top-1 ranking under both segment-level and patient-level CV, and Top-3 ranking under the two patient-centered regimes. The Top-3 feature set {Pressure_peak, WSS_peak/OSI, Vorticity_MSD} was preserved across all four CV regimes (3 of 3 set overlap; Top-5 agreement: 4 of 5), with pairwise Spearman rank correlations spanning ρ=+0.845 to +0.982 (all P<1.1×10⁻³); the mean- vs. median-centered sensitivity yielded ρ=+0.982 (P=8.4×10⁻⁸). These results indicate that mechanistic driver identification is robust to CV specification even where the cross-patient predictive R² collapses, supporting the methodological separation of within-lesion mechanistic inference from cross-patient predictive generalization adopted in this study.

Finally, to validate the necessity of an ensemble approach over linear formulations, the feature interaction structure was analyzed (Figure 6C). The correlation matrix revealed complex physiological coupling among hemodynamic parameters. The engineered WSS_peak/OSI ratio exhibited moderate correlation with canonical WSS_peak (r=0.64); yet its distinct high ranking in the SHAP analysis (Table 7) underscores the Random Forest model’s ability to isolate non-linear predictive signals not captured by linear definitions. Strong coupling among shear-derived metrics further confirms the relevance of an algorithm that does not require strict orthogonality of inputs; the Random Forest navigates these non-linear feature interactions to isolate the specific, independent contribution of local Pressure_peak to wall remodeling. Cross-patient generalizability was then evaluated under strict patient-level Stratified GroupKFold CV; at this cohort size (n=24), cross-patient prediction of absolute EWT was not supported (Figure 5; full per-patient breakdown in Tables S15,S16). This dissociation between the robust within-lesion segment-level fit and the limited cross-patient transferability is consistent with the reduced cross-patient predictive performance noted above.

Figure 5 Patient-level stratified GroupKFold: cross-patient generalization disclosure. (A) Predicted vs. actual EWT under patient-level CV. Scatter density of segment-level random forest predictions evaluated under strict patient-level stratified GroupKFold cross-validation (24 patients held out per fold; n=12,473 segments from 13 AWE-positive patients shown). Each segment is color-coded by anonymized patient identifier (P01–P13). The diagonal indicates perfect prediction (y = x). Predictions cluster horizontally around the global mean rather than tracking the diagonal, yielding a negative coefficient of determination and an overall mean absolute error larger than the inter-patient standard deviation of mean EWT, which together disclose that absolute cross-patient prediction of EWT magnitude is not supported at the current cohort size (quantitative values in Table 6 and Tables S13,S14). (B) Per-patient MAE breakdown. Mean absolute error in EWT prediction for each of the 13 AWE-positive patients, sorted by MAE in descending order. Bars are colored consistently with panel (A) for visual correspondence. The horizontal dashed reference line (blue) marks the inter-patient σ_between =0.1299 mm, and the dotted reference line (red) marks the overall MAE =0.2377 mm. Seven patients (P09, P01, P02, P05, P12, P11, P03) exhibit MAE values exceeding σ_between, indicating that for these individuals the model error surpasses the natural between-patient variability in mean EWT. The number of contributing segments per patient (n shown inside each bar) is not monotonically related to MAE, indicating that prediction limitation is not driven by per-patient segment count. This patient-level disclosure complements the within-lesion co-primary fit reported in Figure 6A, and supports the methodological separation of within-lesion mechanistic inference from cross-patient predictive generalization adopted throughout this work. AWE, aneurysmal wall enhancement; CV, cross-validation; EWT, enhanced wall thickness; MAE, mean absolute error.

Discussion

While the PHASES score remains the clinical standard, its heavy weighting of size may lead to an underestimation of risk in small aneurysms. A recent multicenter cohort reported that nearly 79% of ruptured cases had a PHASES score <5 (18), demonstrating that a preponderance of aneurysmal ruptures occurs in lesions previously stratified as low-risk by the PHASES and UIATS frameworks. This highlights a critical diagnostic gap where small but high-risk lesions escape detection by traditional stratification metrics. Indeed, longitudinal multicenter data confirm that AWE is an independent marker of aneurysm instability that complements traditional diameter-based metrics (19). This supports the hypothesis that biological transformation within the aneurysm wall occurs prior to gross structural changes, indicating that AWE may identify high-risk lesions before they meet conventional size-based thresholds for intervention (20,21). Our study addresses these diagnostic limitations by integrating CFD with high-resolution quantitative vessel wall characterization. We demonstrate that aneurysm instability is not a stochastic occurrence but the manifestation of a distinct mechanobiological coupling. In this framework, localized hemodynamic vectors act as the primary drivers that dictate the spatial distribution and progression of pathological wall remodeling. Regarding baseline characteristics, our cohort confirmed that irregular morphologies (higher aspect/size ratios) and specific ACoA configurations (e.g., A1 dominance) are associated with AWE. However, rather than viewing these as independent geometric risk factors, we interpret them as structural precursors that shape the hemodynamic environment; complex geometries inherently predispose the aneurysm to the flow stagnation identified in our subsequent analysis. Notably, systemic risk factors were largely non-contributory; the isolated statistical significance of alcohol consumption is likely a stochastic artifact of the limited sample size rather than a biological signal.

Our hemodynamic profiling revealed that AWE is not a global mural phenomenon but is site-specifically localized to hemodynamic niches of low WSS and pressure. These regions represent zones of hemodynamic stagnation, where the prolonged contact between inflammatory cells and the aneurysm wall catalyzes the enhancement observed on MRI. This high-resolution mapping indicates that local flow stasis is a primary determinant of the spatial pattern of aneurysm wall vulnerability. This striking spatial concordance provides robust empirical support for a stagnation-driven inflammatory infiltration mechanism. In these low-energy recirculation zones, the prolonged residence time of circulating leukocytes and pro-inflammatory cytokines facilitates their sustained interaction with the endothelium, ultimately catalyzing their transmural infiltration (22,23). This stagnation-driven mechanism is statistically substantiated by our GEE analysis (cluster-robust by PatientID), where low Pressure_peak emerged as the dominant independent association with enhancement presence (β=−0.5452, robust P=0.0499). Low WSS_peak shows a directionally consistent but non-significant association under the same model (β=−0.131, robust P=0.26), reflecting the collinearity between TAWSS and WSS_peak that becomes apparent after proper variance adjustment. This quantitative evidence aligns with the low-shear inflammatory hypothesis proposed by Meng et al. (24), confirming that a specific hemodynamic environment is required to upregulate inflammatory adhesion molecules on the endothelium, a fundamental response well-established in vascular mechanobiology (25). Furthermore, our intra-patient comparison confirms that this hemodynamic deprivation is both profound and persistent. Specifically, enhanced regions exhibited a drastic reduction in WSS_peak (mean difference =−3.97 Pa; P_FDR <0.001) compared to healthy walls, indicating a loss of physiological scouring force. Crucially, this stagnation persists throughout the cardiac cycle. Although TAWSS exhibited a marginal trend following strict FDR correction (mean difference =−0.79 Pa; P_FDR =0.061), its directional alignment with the profound peak-systolic WSS deficit reinforces the phenotype of continuous flow stasis. Collectively, these findings indicate that AWE in our cohort is associated with low-shear stagnation. While high shear stress has also been implicated in wall remodeling, our results align with recent phenotypic classifications, which identifies low WSS as the specific determinant of the thick-walled, developed aneurysms observed in this study. This consistency supports the validity of our hemodynamic calculation, suggesting that the enhancement captured here reflects a stagnation-driven inflammatory pathway (21,26).

Distinct from the initiation phase, our results indicate that the severity of wall pathology is governed by a composite hemodynamic profile. In the segment-level Pearson correlation analysis, TA-Viscosity exhibited the strongest correlation with both anatomical thickening (EWT, r=0.217) and inflammatory intensity (MeanCR, r=0.244) (Tables S12,S13). However, the corresponding intra-patient paired comparison was only marginally trending and did not survive Benjamini-Hochberg correction (Mean Diff =+7.0×10⁻⁴ Pa·s, P_FDR =0.068; Table 2), suggesting that the segment-level Pearson signal partly reflects within-patient correlation structure rather than an independent driver effect. This inconsistency between univariate correlations and the hierarchical paired analysis motivates the cluster-aware approach taken below. Crucially, our comprehensive multi-modal analysis identified Pressure_peak as the dominant independent predictor. In the multivariate LMM, Pressure_peak (β=−0.181, P=2.31×10⁻¹¹) far exceeded the effect of viscosity (β=−0.020, P=1.5×10⁻³), WSSG_peak (β=−0.016, P=1.7×10⁻³), and the small positive coefficients of AFI_MSD (β=+0.021, P=1.8×10⁻⁷) and WSS_peak (β=+0.025, P=4.2×10⁻⁶; the latter became non-significant at caP =10,000, identifying it as a collinearity-driven artefact rather than an independent biological signal). This finding was robustly corroborated by the non-linear Random Forest algorithm, where Pressure_peak ranked as the most influential feature (SHAP Rank 1; Mean |SHAP| =0.0301), outperforming all shear-derived metrics (Table 7). The independent nature of this pressure effect is validated by the feature correlation matrix (Figure 6C), which shows that Pressure_peak possesses a distinct low-correlation profile compared to the highly collinear shear-viscosity metrics. Within these low-pressure stagnation zones, chronic shear deprivation downregulates endothelial protective transcription factors (KLF2/4) (27,28), compromising barrier function and manifesting radiologically as elevated endothelial permeability (K_trans), which has been quantitatively correlated with the severity of wall enhancement in recent clinical cohorts (29). Computational studies have similarly identified rupture-prone aneurysm regions, particularly blebs, as zones of coupled flow stagnation and diminished wall tension (13,30); biomechanical features that align directly with the low Pressure_peak environment quantified in our cohort.

Across our hierarchical analyses, the presence and the severity of wall enhancement converged on a common dominant hemodynamic driver. The GEE analysis on the full segment cohort (predicting AWE presence) and the multivariate LMM on the enhanced subset (predicting EWT magnitude) both identified Pressure_peak as the strongest independent association (GEE β=−0.5452, robust P=0.0499; LMM β=−0.181, P=2.31×10⁻¹¹). This convergence, achieved across analytical strategies that handle within-patient clustering by mathematically distinct mechanisms (cluster-robust sandwich estimators vs. random-intercept hierarchical modeling), substantially strengthens the interpretation that local hypotension, rather than shear stress, is the primary independent hemodynamic determinant of both the spatial localization and the magnitude of wall enhancement in our cohort. Our use of two complementary outcome models reflects the zero-inflated structure of the EWT data rather than a hypothesis of distinct drivers: because healthy wall segments have an EWT of 0 mm, including them in the severity analysis would bias the linear estimation. We therefore used the GEE on the full segment dataset (n=157,284) to identify predictors of AWE presence, and restricted the LMM and machine-learning analyses to enhanced segments (n=12,473, EWT >0) to quantify drivers of pathological thickening among lesional tissue.

A critical question regarding the biological substrate of the quantified EWT is whether it represents nonspecific inflammation or a distinct structural phenotype. While traditional histological studies often failed to resolve the microvascular architecture within aneurysm walls, recent high-resolution 3D imaging of a surgically harvested specimen by Abdurakhmonov et al. offers new insights (31). Their analysis of an intact aneurysm dome revealed that the thickened wall was densely populated by an interconnected vasa vasorum plexus, documenting a high quantitative correlation (R²=0.97) between wall thickness and neovascular density in this index case. This suggests that the enhancement on MRI may reflect this neovascular burden. This finding aligns with the transport modeling by Cebral et al. (32), which indicates that wall thickening in stagnation zones leads to local hypoxia. Our results provide statistical support for this mechanism: the negative association between Pressure_peak and EWT (β=−0.181) confirms that thickening occurs in low-pressure, perfusion-deficient zones.

To capture the spatial heterogeneity inherent in aneurysm pathology, our study utilized the VISP pipeline to map hemodynamic variables directly to the centroid of each triangular surface facet via barycentric interpolation. While global averaging approaches have provided valuable insights, they carry the inherent risk of diluting focal pathological signals (33). Our results confirm the necessity of this high-fidelity surface mapping through the striking consistency of our findings across all statistical domains: focal Pressure_peak emerged as the distinguishing feature in intra-patient paired analysis (Mean Diff =−35.25 Pa, P_FDR =0.02), the dominant independent association under cluster-robust GEE (β=−0.5452, robust P=0.0499), the strongest independent driver in the multivariate LMM with random intercepts (β=−0.181, P=2.31×10⁻¹¹; ICC =0.669), and the top-ranked feature in SHAP analysis (Rank 1; preserved within the Top-3 set across all four CV regimes). These results suggest that a global mean approach might inadvertently homogenize these discrete hypotension zones, potentially obscuring the primary driver of wall remodeling.

Furthermore, the complexity of the hemodynamic-pathological coupling necessitated the use of ensemble machine learning to capture non-linear parameter interactions. As detailed in the feature importance ranking (Table 7), simple linear definitions often fail to characterize the multifaceted mechanical environment sensed by the vessel wall. Our Random Forest model revealed that vulnerability is governed by composite hemodynamic stress profiles rather than isolated variables. Notably, the engineered WSS_peak/OSI ratio (Rank 2; Mean |SHAP| =0.0291), which couples shear magnitude with directional instability, exceeded the predictive contribution of canonical WSS_peak alone (Rank 5; Mean |SHAP| =0.0112). This indicates that the signal for wall thickening lies in coupled rather than isolated shear metrics. The Random Forest model thereby identifies composite hemodynamic niches, such as regions where low Pressure_peak coincides with elevated oscillatory shear, that univariate analyses inherently miss.

A complementary methodological consideration concerns the unit of analysis. Our high-resolution surface mapping treats individual wall segments as the locus of mechanobiological inference, predicated on the hypothesis that AWE represents a localized response governed by local physical conditions rather than a uniform systemic attribute. The observation that AWE is frequently heterogeneous, affecting specific wall segments while sparing others within the same aneurysm dome, underscores this premise: were systemic factors the sole determinants, enhancement would manifest globally rather than at discrete loci. Within-segment dependence is formally addressed by the cluster-aware methods detailed above (LMM random intercepts and GEE cluster-robust SEs), while the pooled segment-level Random Forest provides a complementary, fully data driven view of within-lesion hemodynamic-EWT coupling.

This methodological architecture should not, however, be conflated with cross-patient predictive generalization. The segment-level R² of 0.78 reflects within-lesion hemodynamic-EWT coupling under intra-cluster CV, whereas the corresponding patient-level Stratified GroupKFold CV indicates that absolute cross-patient prediction of EWT magnitude is not supported at this cohort size, reflecting the limited number of patients (full quantitative metrics in Table 6 and Tables S15,S16). Critically, the mechanistic driver hierarchy, with Pressure_peak dominant across all four CV regimes and the Top-3 feature set fully preserved, is itself robust to this distinction, supporting the separation of within-lesion mechanistic inference from cross-patient predictive generalization adopted throughout this work.

Beyond this segment-level vs. cross-patient distinction, our analytical framework cross-validates the inferential conclusion across four hierarchical methods that handle within-patient clustering by mathematically distinct mechanisms: intra-patient paired bootstrap (n=13 patients), multivariate LMM with random intercepts on cluster-balanced lesional segments, GEE with cluster-robust SEs on the full segment cohort (n=157,284), and SHAP feature importance across four CV regimes. The convergence of these methods on Pressure_peak as the dominant independent driver, with no method identifying a contradicting hierarchy, supports the interpretation that this hemodynamic fingerprint represents a robust biological signal rather than a statistical artifact of within-patient dependence (i.e., pseudoreplication).

Finally, we prioritized objective quantification of EWT over subjective visual grading. Consistent with the consensus that quantitative metrics (e.g., the wall enhancement index) offer superior reproducibility compared to binary classification (34,35), our workflow integrated rigorous sub-voxel sampling. While 7T MRI offers superior resolution, recent validation studies (36) demonstrate that wall enhancement patterns strongly correlate with true anatomical thickening, ensuring the biological robustness of using enhancement as a surrogate marker. Importantly, our framework demonstrates that point-wise hemodynamic-wall mapping is feasible on the 3T systems used in routine clinical practice, rather than being confined to specialized research-grade 7T platforms. The agreement between our results and established pathological models indicates that the sub-voxel VISP sampling strategy recovers robust biological signals at this field strength, underscoring the clinical applicability of the approach.

We propose this integrated mechanobiological framework as a basis for non-invasive risk stratification. By identifying focal regions of inflammatory remodeling derived from hemodynamic signatures, this approach offers a potential surrogate for identifying unstable vessel walls beyond simple contrast retention. This distinction has direct relevance for therapeutic decision-making. Recent radiomic analyses indicate that quantitative vessel wall features outperform morphological metrics in predicting rupture risk (37), and pre-treatment enhancement has been specifically linked to higher recurrence rates following standard endovascular coiling (38). Consequently, quantifying the burden of wall instability (EWT) aids in identifying high-risk patients who may require definitive flow reconstruction. For lesions exhibiting a high EWT burden, flow diversion may offer a theoretical advantage over simple coiling. By reconstructing the parent artery and promoting endothelialization, flow diverters aim to isolate the aneurysm wall from the local hemodynamic stagnation that drives the hypoxic-angiogenic cycle identified by our findings (39).

Importantly, our study highlights the value of high-resolution spatial mapping as a complement to patient-level binary diagnosis. Aneurysm rupture is typically precipitated by focal structural failure at a specific weak point (e.g., a daughter sac) rather than global instability (40). Our segment-level analysis pinpoints localized zones of vulnerability, specifically loci where low Pressure_peak coincides with wall thickening, providing a more granular risk landscape than whole-aneurysm classification alone. This point-wise precision may aid neurointerventionists in targeting the specific locus of instability, potentially informing more tailored intervention strategies.

Limitations

This study has several limitations. Primarily, the retrospective, single-center design inherently introduces selection bias, and the cohort of n=24 patients constrains the statistical power for baseline clinical associations. In particular, the statistically significant difference in alcohol consumption between groups (Table 1; P=0.031) reflects a perfect-separation pattern: all four alcohol-positive patients in our cohort fell into the AWE-negative subgroup, with none in the AWE-positive subgroup. This configuration is most parsimoniously interpreted as a stochastic artefact of the limited sample size rather than a biological signal, and the alcohol-related observation should be regarded as hypothesis-generating only.

Cross-patient predictive generalization was not an aim of this study; the ensemble analyses were positioned as mechanistic throughout. The segment-level Random Forest (R²=0.78) indexes within-lesion hemodynamic-EWT coupling and is retained as a co-primary mechanistic inference, not as a generalization claim. Under strict patient-level Stratified GroupKFold CV, absolute cross-patient prediction of EWT magnitude could not be established at the present cohort size (n=24). With so few patients, the held-out cross-patient error is itself estimated with substantial variance, so this result reflects a cohort too small to support, or to reliably evaluate, a cross-patient predictive model, rather than a demonstrated ceiling on the underlying coupling (quantitative metrics in Table 6 and Tables S15,S16). Importantly, the mechanistic conclusion does not depend on cross-patient prediction: the dominant-driver hierarchy led by Pressure_peak is preserved across all four CV regimes, including the patient-level split, even where the predictive R² collapses. Cohort scaling, patient-specific 4D flow MRI, and prospective external validation will be required before this coupling can be evaluated as a cross-patient prediction tool.

Regarding the computational simulation, boundary conditions were prescribed from a population-average 1D vascular tree database (63 major arterial segments) rather than from patient-specific phase-contrast MRI. Published 4D and phase-contrast MRI data in the posterior circulation report inter-subject coefficient of variation ≈15–25% for basilar and vertebral mean flow and ≈10–15% for systolic peak velocity, translating to approximately 20–30% uncertainty in the absolute magnitude of computed WSS and intraluminal pressure. The reported group-mean absolute magnitudes (e.g., ∆Pressure_peak =−35.25 Pa) should therefore be interpreted as carrying this boundary-condition scaling uncertainty in their absolute value. Critically, the intra-patient paired design used in all primary analyses cancels any systematic scaling error by construction, since any non-patient-specific scaling affects both enhanced and non-enhanced segments of the same patient equally; what is preserved by the paired design is the relative spatial topology of the hemodynamic fields, precisely the quantity used to formulate our mechanistic claims. The CFD simulations also utilized a rigid-wall assumption; while fluid-structure interaction models quantify wall tension more directly (41), recent validation studies confirm that rigid-wall CFD accurately captures the spatial topology of flow stagnation zones (42), and the relative distribution of the low-shear regions driving our analysis remains consistent under this assumption (43).

In terms of image acquisition, the sub-voxel resolution of the VISP strategy reflects adaptive interpolation rather than native acquisition resolution; absolute EWT magnitudes therefore remain bounded by the 0.6 mm voxel and are most reliable when interpreted on the relative, within-lesion basis used throughout our analyses. Finally, direct histological correlation was not available for this specific cohort, and future prospective studies with matched histological validation are warranted.

Conclusions

In ACoA aneurysms, we coupled patient-specific CFD with quantitative 3T VWI at the facet level to characterize the local mechanical environment associated with wall enhancement. Across four independent, hierarchical analyses (intra-patient paired bootstrap testing, a LMM, GEE, and SHAP-interpreted ensemble learning), local Pressure_peak emerged consistently as the dominant independent hemodynamic driver of wall thickening in enhanced segments, complementing the established low-WSS association within stagnation zones. By resolving these associations at sub-voxel sampling density on a routine clinical platform, our facet-level approach localizes focal hemodynamic determinants of wall enhancement that whole-aneurysm averaging tends to obscure, without recourse to research-grade 7T imaging. We therefore advance this framework as a mechanistic, explanatory account of local hemodynamic-AWE coupling; its clinical translation will require prospective, externally validated study in larger, multicenter cohorts.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0249/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0249/dss

Funding: This work was supported by the National Natural Science Foundation of China (grant Nos. 82571691 and 82201427 to X.F.), the Young Scientists Fund of the Natural Science Foundation of Sichuan Province (grant No. 2024NSFSC1705 to Y.M.), the Fund for Fostering Distinguished Young Scholars of North Sichuan Medical College (grant No. CBY22-JQ02 to Y.M.), Provincial Undergraduate Training Program on Innovation and Entrepreneurship of Sichuan Province (grant No. S202410634088 to Y.F.), and the Special Funds for the Cultivation of Guangdong College Students’ Scientific and Technological Innovation (grant No. pdjh2025bk052 to C.H.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0249/coif). The 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 the Ethics Committee of Zhujiang Hospital, Southern Medical University (IRB number: 2023-KY-038-03) and individual consent for this retrospective analysis was waived.

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. Vlak MH, Algra A, Brandenburg R, Rinkel GJ. Prevalence of unruptured intracranial aneurysms, with emphasis on sex, age, comorbidity, country, and time period: a systematic review and meta-analysis. Lancet Neurol 2011;10:626-36. [Crossref] [PubMed]
2. Neifert SN, Chapman EK, Martini ML, Shuman WH, Schupper AJ, Oermann EK, Mocco J, Macdonald RL. Aneurysmal Subarachnoid Hemorrhage: the Last Decade. Transl Stroke Res 2021;12:428-46. [Crossref] [PubMed]
3. Frączek MJ, Krzyżewski RM, Kliś KM, Kwinta BM, Popiela TJ, Stachura K. Unruptured intracranial aneurysms: Why should we focus on small aneurysms? A comprehensive update of recent findings. Pol J Radiol 2024;89:e13-23.
4. Krystkiewicz K, Kowal A, Krystkiewicz-Orzechowska M, Arczewski F, Dziedzic K, Tosik M. Unruptured Intracranial Aneurysm Risk Scores Underperform in Predicting Subsequent Rupture: A Retrospective Single-Center Study. Neurol Int 2025;17:189. [Crossref] [PubMed]
5. Larson AS, Lehman VT, Lanzino G, Brinjikji W. Lack of Baseline Intracranial Aneurysm Wall Enhancement Predicts Future Stability: A Systematic Review and Meta-Analysis of Longitudinal Studies. AJNR Am J Neuroradiol 2020;41:1606-10. [Crossref] [PubMed]
6. Digpal R, Arkill KP, Doherty R, Yates J, Milne LK, Broomes N, Katsamenis OL, Macdonald J, Ditchfield A, Narata AP, Darekar A, Carare RO, Fabian M, Galea I, Bulters D. A Systematic Review and Meta-Analysis of the Pathology Underlying Aneurysm Enhancement on Vessel Wall Imaging. Int J Mol Sci 2024;25:2700. [Crossref] [PubMed]
7. Lu Y, Wang C, Bao Y, Huang L, Lu G, Li Y. Association between intracranial aneurysm wall enhancement and intracranial atherosclerotic plaque: a cross-sectional study using high-resolution vessel wall imaging. Quant Imaging Med Surg 2024;14:1553-63. [Crossref] [PubMed]
8. Lauric A, Silveira L, Lesha E, Breton JM, Malek AM. Aneurysm presence at the anterior communicating artery bifurcation is associated with caliber tapering of the A1 segment. J Neurosurg 2022;136:1694-704. [Crossref] [PubMed]
9. Guo H, Liu JF, Li CH, Wang JW, Li H, Gao BL. Greater hemodynamic stresses initiate aneurysms on major cerebral arterial bifurcations. Front Neurol 2023;14:1265484. [Crossref] [PubMed]
10. Li YH, Liu B, Liang ZL, Yang YM, Wang SL, Gao B. Internal carotid artery bifurcation aneurysms are probably formed to decrease the abnormally-enhanced hemodynamic stresses caused by direct flow impaction. Quant Imaging Med Surg 2025;15:10936-49. [Crossref] [PubMed]
11. Tian Y, Li X, Zhang J, Zhao B, Liang F. Identifying hemodynamic factors associated with the rupture of anterior communicating artery aneurysms based on global modeling of blood flow in the cerebral artery network. Front Bioeng Biotechnol 2024;12:1419519. [Crossref] [PubMed]
12. Hadad S, Mut F, Chung BJ, Roa JA, Robertson AM, Hasan DM, Samaniego EA, Cebral JR. Regional Aneurysm Wall Enhancement is Affected by Local Hemodynamics: A 7T MRI Study. AJNR Am J Neuroradiol 2021;42:464-70. [Crossref] [PubMed]
13. Galloy AE, Raghuram A, Nino MA, Varon Miller A, Sabotin R, Osorno-Cruz C, Samaniego EA, Raghavan SML, Hasan D. Analysis of Cerebral Aneurysm Wall Tension and Enhancement Using Finite Element Analysis and High-Resolution Vessel Wall Imaging. Front Neurol 2021;12:764063. [Crossref] [PubMed]
14. Veeturi SS, Rajabzadeh-Oghaz H, Pintér NK, Waqas M, Hasan DM, Snyder KV, Siddiqui AH, Tutino VM. Aneurysm risk metrics and hemodynamics are associated with greater vessel wall enhancement in intracranial aneurysms. R Soc Open Sci 2021;8:211119.
15. Mei Y, Guan D, Tong X, Liu Q, Hu M, Chen G, Li C. Association of cerebral infarction with vertebral arterial fenestration using non-Newtonian hemodynamic evaluation. Math Biosci Eng 2022;19:7076-90. [Crossref] [PubMed]
16. Bernsdorf J, Wang D. Non-Newtonian blood flow simulation in cerebral aneurysms. Comput Math Appl 2009;58:1024-9.
17. Xiang J, Tutino VM, Snyder KV, Meng H. CFD: computational fluid dynamics or confounding factor dissemination? The role of hemodynamics in intracranial aneurysm rupture risk assessment. AJNR Am J Neuroradiol 2014;35:1849-57.
18. Pettersson SD, Skrzypkowska P, Pietrzak K, Och A, Siedlecki K, Czapla-Iskrzycka A, Klepinowski T, Fodor T, Filo J, Meyer-Szary J, Fercho J, Sunesson F, Olofsson HKL, Ali S, Szmuda T, Miekisiak G. Evaluation of PHASES Score for Predicting Rupture of Intracranial Aneurysms: Significance of Aneurysm Size. World Neurosurg 2024;184:e178-84. [Crossref] [PubMed]
19. van der Kamp LT, Edjlali M, Naggara O, Matsushige T, Bulters DO, Digpal R, Zhu C, Saloner D, Hu P, Zhai X, Mossa-Basha M, Tian B, Sakamoto S, Fu Q, Ruigrok YM, Zhao H, Chen H, Rinkel GJE, van der Schaaf IC, Vergouwen MDI. Gadolinium-enhanced intracranial aneurysm wall imaging and risk of aneurysm growth and rupture: a multicentre longitudinal cohort study. Eur Radiol 2024;34:4610-8. [Crossref] [PubMed]
20. Frösen J, Cebral J, Robertson AM, Aoki T. Flow-induced, inflammation-mediated arterial wall remodeling in the formation and progression of intracranial aneurysms. Neurosurg Focus 2019;47:E21. [Crossref] [PubMed]
21. Tang Y, Wei H, Zhang Z, Fu M, Feng J, Li Z, Liu X, Wu Y, Zhang J, You W, Xue R, Zhuo Y, Jiang Y, Li Y, Li R, Liu P. Transition of intracranial aneurysmal wall enhancement from high to low wall shear stress mediation with size increase: A hemodynamic study based on 7T magnetic resonance imaging. Heliyon 2024;10:e30006. [Crossref] [PubMed]
22. Zhong W, Du Y, Kuang H, Liu M, Xue F, Bai X, Wang D, Su W, Wang Y. Hemodynamic Characteristic Analysis of Aneurysm Wall Enhancement in Unruptured Middle Cerebral Artery Aneurysm. Front Neurol 2022;13:781240. [Crossref] [PubMed]
23. Shantha Kumar V, Shantha Kumar V. High Wall Shear Incites Cerebral Aneurysm Formation and Low Wall Shear Stress Propagates Cerebral Aneurysm Growth. J Neurol Res 2023;13:1-11.
24. Meng H, Tutino VM, Xiang J, Siddiqui A. High WSS or low WSS? Complex interactions of hemodynamics with intracranial aneurysm initiation, growth, and rupture: toward a unifying hypothesis. AJNR Am J Neuroradiol 2014;35:1254-62.
25. Souilhol C, Serbanovic-Canic J, Fragiadaki M, Chico TJ, Ridger V, Roddie H, Evans PC. Endothelial responses to shear stress in atherosclerosis: a novel role for developmental genes. Nat Rev Cardiol 2020;17:52-63. [Crossref] [PubMed]
26. Karnam Y, Mut F, Yu AK, Cheng B, Amin-Hanjani S, van Keulen M, Charbel FT, White T, Niemelä M, Tulamo R, Rezai Jahromi B, Frösen J, Tobe Y, Robertson AM, Cebral JR. Intracranial Aneurysm Wall Phenotypes: Clinical, Morphological, and Hemodynamic Influences. Stroke Vasc Interv Neurol 2025;5:e001829. [Crossref] [PubMed]
27. Joshi D, Coon BG, Chakraborty R, Deng H, Yang Z, Babar MU, Fernandez-Tussy P, Meredith E, Attanasio J, Joshi N, Traylor JG Jr, Orr AW, Fernandez-Hernando C, Libreros S, Schwartz MA. Endothelial γ-protocadherins inhibit KLF2 and KLF4 to promote atherosclerosis. Nat Cardiovasc Res 2024;3:1035-48. [Crossref] [PubMed]
28. Zheng Q, Zou Y, Teng P, Chen Z, Wu Y, Dai X, Li X, Hu Z, Wu S, Xu Y, Zou W, Song H, Ma L. Mechanosensitive Channel PIEZO1 Senses Shear Force to Induce KLF2/4 Expression via CaMKII/MEKK3/ERK5 Axis in Endothelial Cells. Cells 2022;11:2191. [Crossref] [PubMed]
29. Fu Q, Zhang Y, Zhang Y, Liu C, Li J, Wang M, Luo H, Zhu J, Qu F, Mossa-Basha M, Guan S, Cheng J, Zhu C. Wall permeability on magnetic resonance imaging is associated with intracranial aneurysm symptoms and wall enhancement. Eur Radiol 2024;34:5204-14. [Crossref] [PubMed]
30. Sanchez S, Miller JM, Galloy A, Nino M, Ojeda D, Sagues E, Dier C, Gudino A, Jaramillo R, Samaniego EA. Comprehensive morphomechanical and wall enhancement analysis of intracranial aneurysms. Eur Radiol 2025;35:5027-36. [Crossref] [PubMed]
31. Abdurakhmonov M, Tobe Y, Cebral JR, Asadbeygi A, Ramezanpour M, Zamani M, Watkins SC, Phillippi JA, Yu AK, Amin-Hanjani S, Charbel FT, Kaneko N, Cheng BC, Dehdashti AR, White T, Robertson AM. Vasa vasorum plexus formation in intracranial aneurysm associates with microcalcification and wall thickening. Sci Adv 2025;11:eadx4374. [Crossref] [PubMed]
32. Cebral JR, Mut F, Lohner R, Abdurakhmonov M, Ramezanpour M, Tobe Y, Robertson AM. Analysis of the Role of Vasa Vasorum in the Oxygen Transport to the Aneurysm Wall. Int J Numer Method Biomed Eng 2025;41:e70051. [Crossref] [PubMed]
33. Fu M, Peng F, Zhang M, Chen S, Niu H, He X, Xu B, Liu A, Li R. Aneurysmal wall enhancement and hemodynamics: pixel-level correlation between spatial distribution. Quant Imaging Med Surg 2022;12:3692-704. [Crossref] [PubMed]
34. Fu Q, Wang Y, Zhang Y, Zhang Y, Guo X, Xu H, Yao Z, Wang M, Levitt MR, Mossa-Basha M, Zhu J, Cheng J, Guan S, Zhu C. Qualitative and Quantitative Wall Enhancement on Magnetic Resonance Imaging Is Associated With Symptoms of Unruptured Intracranial Aneurysms. Stroke 2021;52:213-22. [Crossref] [PubMed]
35. Veeturi SS, Pinter NK, Monteiro A, Baig AA, Rai HH, Waqas M, Siddiqui AH, Rajabzadeh-Oghaz H, Tutino VM. An Image-Based Workflow for Objective Vessel Wall Enhancement Quantification in Intracranial Aneurysms. Diagnostics (Basel) 2021;11:1742. [Crossref] [PubMed]
36. Liu X, Feng J, Li Z, Zhang Z, Zhang Q, Jiang Y, Huo X, Chai X, Wu Y, Kong Q, Liu P, Ge H, Jin H, An J, Jiang P, Saloner DA, Li Y, Zhu C. Quantitative analysis of unruptured intracranial aneurysm wall thickness and enhancement using 7T high resolution, black blood magnetic resonance imaging. J Neurointerv Surg 2022;14:723-8. [Crossref] [PubMed]
37. Yuan W, Jiang S, Wang Z, Yan C, Jiang Y, Guo D, Chen T. High-resolution vessel wall imaging-driven radiomic analysis for the precision prediction of intracranial aneurysm rupture risk: a promising approach. Front Neurosci 2025;19:1581373. [Crossref] [PubMed]
38. Ladenhauf V, Galijasevic M, Regodic M, Helbok R, Rass V, Freyschlag C, Petr O, Deeg J, Gruber L, Mangesius S, Gizewski ER, Grams AE. Aneurysmal Wall Enhancement of Non-Ruptured Intracranial Aneurysms after Endovascular Treatment Correlates with Higher Aneurysm Reperfusion Rates, but Only in Large Aneurysms. Diagnostics (Basel) 2024;14:1533. [Crossref] [PubMed]
39. Gaub M, Murtha G, Lafuente M, Webb M, Luo A, Birnbaum LA, Mascitelli JR, Al Saiegh F. Flow Diversion for Endovascular Treatment of Intracranial Aneurysms: Past, Present, and Future Directions. J Clin Med 2024;13:4167. [Crossref] [PubMed]
40. Omodaka S, Sugiyama SI, Sakata H, Funamoto K, Yamaguchi T, Niizuma K, Endo H. Aneurysm Wall Enhancement Can Predict Rupture Point in Intracranial Aneurysms With Multiple Blebs. Neurosurgery 2025;96:593-9. [Crossref] [PubMed]
41. Wang H, Balzani D, Vedula V, Uhlmann K, Varnik F. On the Potential Self-Amplification of Aneurysms Due to Tissue Degradation and Blood Flow Revealed From FSI Simulations. Front Physiol 2021;12:785780. [Crossref] [PubMed]
42. Hadad S, Karnam Y, Mut F, Lohner R, Robertson AM, Kaneko N, Cebral JR. Computational fluid dynamics-based virtual angiograms for the detection of flow stagnation in intracranial aneurysms. Int J Numer Method Biomed Eng 2023;39:e3740. [Crossref] [PubMed]
43. Goetz A, Jeken-Rico P, Pelissier U, Chau Y, Sédat J, Hachem E. AnXplore: a comprehensive fluid-structure interaction study of 101 intracranial aneurysms. Front Bioeng Biotechnol 2024;12:1433811. [Crossref] [PubMed]