In-vivo quantitative T₂ mapping of carotid arteries in atherosclerotic patients: segmentation and T₂ measurement of plaque components

The carotid arteries of 2 healthy volunteers and 15 patients with known atherosclerosis (11 males and 4 females, 71 ± 10 years, range 54-84) were imaged on two 3T scanners (12 patients on TIM Trio and 3 on Verio, Siemens Healthcare, both running VB17 software) using the Multiple-Spin-Echo (Multi-SE) sequence, called Spin-Echo-Multi-Contrast (SE_MC) on Siemens systems. Ethics approval from the local board was obtained and all subjects gave informed written consent. Black-blood cross-sectional 2D images of carotid arteries were acquired with the Machnet 4-channel phased-array carotid coil using Double-Inversion-Recovery (DIR) preparation and cardiac gating. Chemical shift selective fat saturation (FAT SAT) was used to suppress the signal from subcutaneous and perivascular fat, mainly composed of triglycerides, without affecting the signal from cholesterol and cholesteryl esters, the main lipids in atherosclerotic plaques [21, 22]. If necessary, a saturation band was positioned on the anterior region of the neck to reduce ghosting artefacts from breathing and swallowing. Low Specific-Absorption-Rate (SAR) pulses were used in order to keep the total dissipated radio-frequency energy below the SAR limit without changing the flip angle of the refocusing pulses. Multi-SE parameters were TE = 12.9, 25.8, 38.7, 51.6, 64.5, 77.4, 90.3, 103.2 ms, TR = 2 R-R intervals, field of view (FOV) = 160 × 128 mm² and matrix size = 320 × 256. Image resolution after zero-padding was 640 × 512 (pixel size = 0.25 mm). Slice thickness (2 mm) and pixel bandwidth (130 Hz/pix) were the same for Multi-SE and FSE. FSE T₁W (TE = 14 ms, TR = 1 R-R and ETL = 9), T₂W (TE = 89 ms, TR = 2 R-R and ETL = 15) and PDW (TE = 14 ms, TR = 2 R-R and ETL = 9) images had FOV = 150 × 150 mm² and matrix size = 320 × 320. Image resolution after zero-padding was 640 × 640 (pixel size = 0.234 mm). TOF angiography was acquired using 3D Fast Low Angle Shot (FLASH) with flip angle = 18°, TR = 45 ms, TE = 4.1 ms, FOV = 200 × 150 mm², matrix size = 256 × 192 and slice thickness = 1 mm. 69 TOF and 11 DIR-FSE T₁W consecutive slices centred on the carotid bifurcation were first imaged in order to localize the atherosclerotic plaque. Then a series of DIR-FSE T₂W and PDW (and if necessary other T₁W) images were acquired to obtain full plaque coverage. Finally, a single-slice DIR-Multi-SE image series was acquired through the plaque centre. Given the constraints imposed by other imaging parameters, the TE series used in Multi-SE was defined considering an expected T₂ ~ 50 ms for the normal carotid wall, which corresponded to a transverse magnetization half-life time T₂·ln(2) ~ 35 ms. The last TE used in this study was approximately equal to 3 half-life times and typical SNR values measured on normal carotid wall went from ≥ 30 at the first echo to ≥ 4 at the last echo. Partial Fourier imaging was used to collect 5/8 of k-space (160 phase-encoding steps) and reduce the acquisition time of single-slice Multi-SE to circa 320 R-R intervals. The quality of Multi-SE and multicontrast images was assessed after each scan by two reviewers (A.C.L. and L.B.) who decided by consensus when carotid wall and plaque were not clearly visible.

The Multi-SE sequence design is similar to the standard FSE that is widely used in carotid imaging, the main difference being the k-space acquisition strategy. Both are Carr-Purcell-Meiboom-Gill (CPMG) based sequences in which a 90° excitation pulse is followed by a train of 180° refocusing pulses to generate multiple spin echoes at every TR [24, 25]. Constant crusher gradients at the left and right of each refocusing slice selective gradient are used to select both primary and stimulated echoes, which sum up at the mid-point between two consecutive refocusing pulses, whereas all other unwanted signal pathways are filtered out. In FSE each echo is phase-encoded to a different k-space line in order to collect multiple k-space lines for a single image during every TR, whereas in Multi-SE the phase-encoding is the same for the entire echo train. In this way, each echo of the Multi-SE sequence fills a line of a series of independent k-space planes, so that a set of images with different TEs are collected (Figure 1). Accuracy and precision improvements for clinical T₂ mapping using a CPMG sequence have been demonstrated by setting the slice thickness of the refocusing pulses larger than that of the excitation pulse [26]. This simple modification stabilizes the flip angle of the refocusing pulse across each slice and minimizes refocusing profile errors. A 1.5 ratio between refocusing and excitation slice thickness was used in this study.

Quantitative T₂ maps of manually drawn regions of interest (ROIs) including vessels and muscles were generated from the Multi-SE series of images. The first image was discarded because of the different nature of the first echo with respect to the others. Since the flip angles of refocusing pulses are not exactly 180° due to B₁ field inhomogeneities, only the signal acquired at the first TE is a pure primary echo, whereas signals acquired at the subsequent TEs are composed by primary and stimulated echoes. The signal intensity (SI) of the first echo will thus not fit in the exponential decay curve followed by the SIs of subsequent echoes, which are enhanced by stimulated echoes. For every ROI voxel the T₂ mono-exponential decay curve SI = β·e ^-TE/T2 was therefore fitted to 7 SIs collected at different TEs (β includes the effect of proton density, T₁ weighting, coil sensitivity, signal amplification and other factors).

The regression model used for T₂ mapping consisted of a nonlinear monoexponential fit (Figure 2) initialized by T₂ and β resulting from a robust linear fit of ln(SI). Parameters estimated by linear least-squares regression are very sensitive to noise because the original error distribution becomes asymmetric after the logarithmic transform. To reduce this sensitivity to outliers, we used a robust regression method that iteratively minimizes the weighted sum of squares using bisquare weights, which depend on the residual of each data point [27]. To assess the statistical significance of fit parameter estimates, the two-tailed p-value of the ratio of each estimate to its standard error (derived from the diagonal elements of the estimated covariance matrix) was calculated on the t-distribution. Only linear fit parameter estimates rejecting the null hypothesis (p < 0.05) were used as initial values for the nonlinear fit in order to improve its convergence. Otherwise default initial values β = 500 and T₂ = 50 ms were used. The nonlinear least-squares regression used the Levenberg-Marquardt algorithm [28]; data points at long TEs with SNR < 2 were discarded; 95% confidence intervals for the exponential curve (Figure 2) and for the fit parameters were estimated. Finally, only voxels with significant T₂ and β estimates (p < 0.05) were accepted for generating T₂ maps. The carotid arteries were subsequently segmented using a semi-automated method [29] that detected inner and outer wall boundaries on the second image (TE = 25.8 ms, TR = 2 R-R) of the Multi-SE series.

After this procedure, T₂ maps of atherosclerotic arteries were segmented into 3 tissue types using a Bayes classifier [30] defined by the probability model P(C _i |T ₂ ) = P(C _i )P(T ₂ |C _i ) combined with the maximum a posteriori (MAP) decision rule. P(C _i |T ₂ ) with i = 1,2,3 is the posterior conditional probability for a voxel with a specific T₂ value to belong to one of 3 classes: LRNC, fibrous tissue or recent IPH. The MAP rule then assigned a particular voxel to the tissue type with the highest posterior probability. P(C _i |T ₂ ) is determined by the prior probability P(C _i ), which was assumed to be equal for every class, and the normal probability distribution P(T ₂ |C _i ), which was estimated from a set of approximately 100 voxels for each class. These voxels were manually selected from all the T₂ maps to represent a specific tissue using multicontrast CMR as a guide, i.e. they were chosen inside a plaque component identified on multicontrast CMR [4, 9]. In order to help matching the corresponding plaque features, multicontrast images were co-registered [31], so that inner and outer vessel boundaries detected on T₁W images [29] could then be superimposed on PDW and T₂W images. Calcification was identified by voxels inside the vessel wall boundaries with SNR < 2 at TE = 14 ms on the T₂ decay curve (synthetic PDW image). Finally, mean ± SD of the T₂ values of voxels classified as LRNC, fibrous tissue and recent IPH were calculated. The algorithms for quantitative T₂ mapping, vessel wall segmentation and plaque classification were implemented in Matlab (Mathworks).

To compare atherosclerotic plaque characterization by T₂ maps with conventional multicontrast CMR, two reviewers (A.C.L. and L.B.) independently classified the carotid arteries of 15 patients as normal or diseased (plaque type III, IV-V, VI, VII or VIII) following the CMR-modified AHA scheme [4, 9]. One reviewer analysed multicontrast CMR (T₁W, T₂W and PDW image intensity inhomogeneity was corrected [13]) and TOF, while the other relied on T₂ maps and TOF at matching slice locations. Both were blinded to the identification and the clinical data of each patient.

The main limitation of this study is the absence of plaque histology, since only two of the participating patients underwent surgery and their endarterectomy specimens were of insufficient quality for detailed analysis. However, plaque characterization performed on in-vivo T₂ maps agreed with results of multicontrast CMR, which has been previously validated by histology [4‐9]. A physical limitation of T₂ mapping is that the presence of fresh IPH (type I) cannot be detected by T₂ alone. Fresh IPH infiltrating the LRNC is characterized by short T₂[8, 9], so it can only be detected using T₁ information, e.g. independent identification on TOF images (Figure 4).

Finally, imperfections of the refocusing pulses represent the main source of errors in T₂ quantitation [33], although the CPMG sequence used in this study partially compensates for them. Stimulated echoes produced by non-ideal refocusing pulses are acquired together with the primary echoes, thus introducing T₁ weighting into the signal and altering the pure T₂ relaxation decay. After discarding the first echo, which is purely primary, the contribution of stimulated echoes to the remaining 7 echoes may have caused T₂ overestimation. These errors depend on the amount of T₁ weighting introduced by the use of TR = 2 R-R. A very long TR (~5·T₁ of target tissue) would limit the T₁ effect, but it would be absolutely impractical under in-vivo high-resolution imaging conditions. Finally, the elimination of the first echo may have affected the estimation of short T₂ components in the LRNC (~35 ms), which have a transverse magnetization half-lifetime (~24 ms) as long as the second TE. In general, this problem can be alleviated by the use of shorter echo spacing (ESP). This study used the shortest ESP possible, given the requirements for matrix size, receiver bandwidth and the duration of low-SAR refocusing pulses. Future developments will aim at improving in-vivo T₂ measurement accuracy, decreasing acquisition time and validating plaque characterization with histology.