In vivo serial MRI data acquisition and segmentation

After informed consent, serial MRI data of carotid atherosclerotic plaques from 16 patients (15 male, 1 female; age: 59–81, mean=71.9; two patients had L/R carotid MRI usable) were acquired 2–4 times (scan interval: 18 months; 4 scans: 7; 3 scans 8; 2 scans: 1) by the University of Washington (UW) Vascular Imaging Laboratory (VIL) using protocols approved by the UW Institutional Review Board. A total of 38 scan pairs (baseline and follow-up) were formed for progression analysis. MRI scans were conducted on a GE SIGNA 1.5-T whole body scanner using an established protocol (Yuan and Kerwin, 2004). Multi-contrast images in T1, T2, proton density (PD), time-of-flight (TOF), and contrast-enhanced (CE) T1 weighted images of atherosclerotic plaques were generated to characterize plaque tissue composition and luminal and vessel wall morphology [14–16]. Segmentation was performed using a custom-designed computer package CASCADE (Computer-Aided System for Cardiovascular Disease Evaluation) developed by the Vascular Imaging Laboratory at the University of Washington. The slice thickness was 2 mm. Field of view = 160 mm × 160 mm. Matrix size 512 × 512 (the real matrix size was 256 × 256. Images were machine interpolated to 512 × 512). After interpolation, the in-plane resolution was 0.31 × 0.31 mm². Figure 1 gives two examples re-constructed from MRI data showing plaque progression and regression, respectively.

3D geometry re-construction and mesh generation

Under in vivo conditions, arteries are axially stretched and pressurized. Therefore, in vivo MRI plaque geometry needs to be shrunk axially and circumferentially a priori to obtain the no-load starting geometry (see Figure 2) for computational simulations as our model starts from the no-load geometry with zero stress/strain, zero pressure and no-flow conditions. The shrinkage in axial direction was 9% so that the vessel would regain its in vivo length with a 10% axial stretch. Circumferential shrinkage for lumen (about 8-12%) and outer wall (about 2-5%)was determined by trial-and-error so that: 1) total mass of the vessel was conserved; 2) the loaded plaque geometry after 10% axial stretch and pressurization had the best match with the original in vivo geometry.

Because advanced plaques have complex irregular geometries and 3D FSI models involve large deformation and large strain, a geometry-fitting mesh generation technique was developed to generate mesh for these models [17, 18]. Using this technique, the 3D plaque and fluid domains were divided into hundreds of small “volumes” to curve-fit the irregular plaque geometries (see Figure 3). Computational meshes for these volumes were then generated automatically by ADINA (ADINA R & D, Inc., Watertown, MA, USA), a commercial finite element software used to solve these FSI models.

3D fluid–structure interaction plaque model and solution methods

where I₁ and I₂ are the first and second strain invariants, C =[C_ij] = X^TX is the right Cauchy-Green deformation tensor, c_i and D_i are material parameters chosen to match experimental measurements and the current literature [21, 22]. Parameter values used in this paper: vessel/fibrous cap, c₁=36.8 KPa, D₁=14.4 KPa, D₂=2; calcification, c₁=368 KPa, D₁=144 KPa, D₂=2.0; lipid-rich necrotic core, c₁=2 KPa, D₁=2 KPa, D₂=1.5; loose matrix, c₁=18.4 KPa, D₁=7.2 KPa; D₂=1.5. c₂ = 0 was set for all materials [17].

Blood flow was assumed to be laminar, Newtonian, viscous and incompressible. The incompressible Navier–Stokes equations with arbitrary Lagrangian–Eulerian (ALE) formulation were used as the governing equations. A no-slip condition, natural traction equilibrium boundary condition and continuity of displacement were assumed on the interface between solid and fluid. Inlet and outlet were fixed (after initial pre-stretch) in the longitudinal (axial) direction, but allowed to expand/contract with flow otherwise. Patient-specific systolic and diastolic pressure conditions from the last hospital admission were used as the maximum and minimum of the imposed pulsatile pressure waveforms at the inlet and outlet of the artery. Details of the FSI model were given in Tang, et al. (2004, 2009) [17, 20].

The 3D FSI models were solved by ADINA, using unstructured finite element methods for both fluid and solid domains. Nonlinear incremental iterative procedures were used to handle fluid–structure interactions. The governing finite element equations for both solid and fluid models were solved by Newton–Raphson iteration method. More details of the computational models and solution methods can be found in Tang et al. (2004, 2009) and Bathe (2002) [18–20]. Plaque wall stress and flow shear stress data corresponding to peak systolic pressure were recorded for analysis.

Plaque progression measurements and data extraction for correlation analysis

Statistical analysis

The Linear Mixed-Effects (LME) models [23] were used to study the correlation between WTI and the predictors (PWS and FSS) at the initial time point of each time pair. The measured points are nodes, and it seems reasonable to capture the dependence among nodes based on their 3-dimensional spatial locations on the vessel. The models are specified as follows.

where y _ij is the WTI value at the i th node on the j th slice, x _ij is the corresponding value of FSS (or PWS). β₀ and β₁ are the fixed effects of the predictor for the baseline and the changing rate of WTI, respectively. The spatial dependence structure among y _ij is accounted for by the exponential isotropic variogram model in 3D space, which is analogous to the autoregressive model in 1D space [23]. Specifically, the vector of random error terms (∈ _ij ) follows a joint Gaussian distribution with mean 0, and the correlation between ${\in }_{i_1}{j_1}_{}$ and ${\in }_{i_2}{j_2}_{}$ is assumed an exponential function ϕ^s, where s is the Euclidean distance between the 3-dimensional spatial locations of the two nodes on the vessel, ϕ is the correlation parameter to be estimated in the model fitting by restricted maximum likelihood (REML) algorithm. The null hypothesis that no correlation exists between WTI and FSS (or PWS) indicates β₁ = 0. A small p-value of the Student’s t-test for the coefficient [23] provides a strong statistical evidence to reject the null and accept the existence of correlation.

where y _ijk is the WTI value at the i th node on the j th slice of the k th patient, x _ijk is the corresponding value of FSS (or PWS). β₀ and β₁ have the same meanings as those in equation (4). For the k th patient, random effect b _k follows a Gaussian distribution with mean 0, which models the clustering dependence of WTI values within this patient. The vector of error terms (∈ _ijk ) follows a joint Gaussian distribution with mean 0. The patients are assumed to be independent so the correlation between the error terms from different patients is 0. The correlation between ${\in }_{i_1}{{j_1}_{}}_k$ and ${\in }_{i_2}{j_{2}k}_{}$ from the same k th patients is assumed an exponential function ${\varphi }_k^s$, where s is the Euclidean distance between the 3D spatial locations, ϕ _k is the patient-specific correlation parameter to be estimated in model fitting. To test the correlation, we again use the p-value for testing β₁ = 0.

To assess the above 3D spatial LME models, we randomly permuted the response WTI values and then calculated the p-values. Such a permutation breaks potential correlations between WTI and FSS (or PWS), so the corresponding p-value is expected to follow a Uniform (0, 1) distribution. Our results verified that the empirical distributions of these p-values after permutations were close to such expectation, which evidenced that our models have the type I error well controlled at the level of individual hypothesis test. That is, these models are appropriate in fitting the correlations between WTI and FSS (or PWS), and the observed small p-values are reliable to indicate the significance of the correlations. We controlled the type I error rate at 0.05.

where ${\widehat{\beta }}_1$ is the estimated slope coefficient by fitting FSS (or PWS) to WTI with the LME model (4) or (5). Because ${\widehat{\beta }}_1$ is estimated in the models that have adjusted the 3D spatial dependence structure among the observations of FSS (or PWS), r is the dependence-adjusted correlation coefficient.

Plaque progression (WTI) correlates positively with plaque wall stress at baseline

Only 16 pairs out of 38 showed negative correlation between WTI and FSS at baseline

Most patients changed correlation sign during the long-term follow-up study

Re-thinking about mechanisms governing advanced plaque progression

A “mechanism” governing a physical or biological process should be something that is true for a majority of events from any given observations, both in terms of samples such as patients and observation time intervals. Then the “mechanism” could be used to predict future trends and possible clinical events, and the prediction should be true at least statistically. There have been huge efforts seeking mechanisms governing plaque progression. It is intuitively natural that low and oscillating flow shear stress would create more favorable flow conditions for cell adhesion and atherosclerosis initiation. Recent advance of medical imaging technology made it possible to track patients non-invasively, obtain patient-specific plaque progression data and quantify possible correlations between plaque progression and various factors such as flow shear stress and plaque wall stress. While the research community has yet to come to a consensus, it is becoming clearer that correlations between advanced plaque progression and mechanical stresses, if they exist, may not be as strong as we thought they were. Results from this study indicated that more than 80% of the patients could not hold their correlation sign over time. If we insist on finding some mechanisms which remained true for all observation intervals for the same patient, those “mechanisms” would be applicable only to 18.75% patients for a positive correlation between plaque progression and plaque wall stress, and 12.5% patients for negative correlation between plaque progression and flow shear stress.

Our results give strong motivations to seeking other mechanisms governing plaque progression, or investigating the reasons causing the correlation sign change. Use of lipid-lowering medications such as statin could be a contributing factor. Diet, cholesterol, mental stress, sudden change of life style, or other disease or illness could all be contributing factors. Growth of plaques depends not only on stress and strain, but a complex biology and physiology environments. Investigation with more possible factors included may lead to new findings.

Limitations

We are limiting our research to the correlation study between plaque progression and mechanical stresses (FSS and PWS). Other risk factors such as stenosis severity, lipid-rich necrotic cores, and intraplaque hemorrhage will be studies in our subsequent investigations. Other model limitations include: a) the use of an isotropic material model for the vessel because patient-specific anisotropic material properties were not available in vivo [24]; b) flow was assumed laminar because the average stenosis severity (by diameter) of the 54 plaques was 50% and laminar flow assumption was considered acceptable at this level [25]; c) arm systole and diastole pressures taken at scan visit were used to scale the pressure profile used in the simulations since pressure conditions right at the location of the plaque were not available; d) effect of statin use and presence of intraplaque hemorrhage (IPH) were not included in the current study due to lack of complete patient data. Development of IPH would be expected to increase FSS on follow-up scans due to progression in luminal narrowing, but result in continued increase in wall thickness. We are continuing our effort and results will be reported as they become available.