INTRODUCTION

MRF-ASL is a recently developed technique that allows simultaneous estimation of CBF, BAT, and tissue T₁ in a single scan of < 4 minutes^1,2. With the use of deep-learning reconstruction methods^3-5, it has demonstrated reliable mapping of hemodynamic parameters. The present work aims to examine the utility of MRF-ASL in the assessment of ischemic stroke. We evaluated MRF-ASL parameters (i.e. CBF, BAT and T₁) in brain regions corresponding to stroke, perilesional tissues, and contralateral normal brain tissues. We also examined the relationship between imaging parameters and the National Institutes of Health Stroke Scale (NIHSS) as well as the modified Rankin Scale (mRS). Finally, we determined the sensitivity of MRF-ASL parameters in differentiating stroke versus normal voxels in the brain.

METHODS

MR experiments:
34 ischemic stroke patients (57.97±13.70 yrs, 14 female) were scanned on a 3T Philips system. The imaging parameters of MRF-ASL: matrix size=64×64×7; voxel size=2.8×2.8×10 mm³; slice gap=1 mm; field-of-view (FoV)=180×180×76 mm³; TE=9.3ms; flip angle 40°; in-plane SENSE factor 2.4; multi-slice EPI; 500 time points; scan duration=3 min 45 sec. A T₂-wighted image was acquired with the following parameters: matrix size=256×256×70; voxel size=0.8×0.8×2.2 mm³; FoV=212×212×154 mm³; TR=4200ms; TE=12ms; scan duration=3 min 28 sec. A diffusion tensor imaging (DTI) sequence was performed with the following parameters: matrix size=256×256×70; voxel size=0.8×0.8×2.2 mm³; FoV=212×212×154 mm³; TR=7000ms; TE=71ms; thirty-three gradient orientations with b=0 and 700 s/mm²; scan duration=4 min 34 sec.
Data processing:
Details of the MRF-ASL processing have been described previously⁵. Briefly, we used an artificial-neural-network (ANN) to reconstruct parametric maps from the raw image series. A total of four parametric maps were obtained from the MRF-ASL scan: CBF_{1-compartment}, CBF_{2-compartment}, BAT, and T₁. Structural images were co-registered to the MRF-ASL space while preserving in-plane resolution. To quantify MRF-ASL parameters in different brain regions, regions-of-interest (ROIs) of ischemic lesion and contralateral normal tissues were manually delineated (Figure 1) by an imaging scientist based on co-registered structural images. A board-certified neuroradiologist also reviewed, adjusted, and confirmed these manually delineated ROIs. DWI was used for ROI drawing in stroke patients within 2 weeks after onset of symptoms (N=10). T_2-weight images were used in patients >=2 weeks after onset (N=24). We also dilated the ischemic lesion ROI by 3.2 mm and referred to these dilated voxels as “perilesional tissue ROI”.
Statistical analysis:
Wilcoxon signed-rank tests were performed to compare MRF-ASL parameters between lesion, perilesional, and normal ROIs. Regression analysis was performed to evaluate the association between NIHSS (or mRS) and MRF-ASL parameters in lesion relative to normal region (i.e. Parameter_lesion−Parameter_normal). A p value (with Bonferroni correction when applicable) of 0.05 or less was considered significant. Receiver-operating-characteristic (ROC) curves were applied to the voxel-wise data to study the sensitivity and specificity of MRF-ASL parameters in correctly classifying voxels within the lesion and normal ROIs.

RESULTS AND DISCUSSION

MRF-ASL parametric maps of four representative stroke patients with a range of NIHSS values are shown in Figure 1. Lesion and contralateral normal ROIs are also displayed.
Figure 2 summarizes MRF-ASL parametric values in ROIs corresponding to stroke lesions, perilesional, and contralateral normal ROIs for CBF_{1-compartment}, CBF_{2-compartment}, BAT, and T₁. Wilcoxon signed-rank tests revealed a significant difference in MRF-ASL parameters between lesion and normal ROIs in all four parameters (p<0.01). Specifically, stroke lesions manifested lower CBF_{1-compartment,} lower CBF_{2-compartment}, longer BAT and longer T₁ compared to normal ROIs. Perilesional parametric values were generally in-between lesion and normal ROI values (p<0.05), except for T₁, where there was no significant difference between perilesional and lesion values.
Table 1 summarizes the associations between MRF-ASL parameters and NIHSS and mRS scores. We found a negative association between acute NIHSS score and CBF_{1-compartment,diff} (p=0.008) and between acute NIHSS score and CBF_{2-compartment,diff} (p=0.003), and a positive association between acute NIHSS score and T_1,diff (p=0.001). That is, a lower CBF_{1-compartment}, lower CBF_{2-compartment}, and longer T₁ are associated with greater neurological deficits at the acute stage. BAT_diff did not show an association with the acute NIHSS score. Figure 3 shows a scatter plot between MRF-ASL parameters and acute NIHSS score. Similar relationships were observed (Table 1) when comparing MRF-ASL parameters with NIHSS and mRS scores acquired on the day of MRI (i.e. non-acute scores). However, the associations with the non-acute scores were generally weaker.
Figure 4A shows group-averaged ROC curves of using MRF-derived CBF_{1-compartment}, CBF_{2-compartment,} BAT, T₁, and their combination in the classification of lesion versus normal voxels. Figure 4B displays the AUC of ROC curves corresponding to these classification schemes. The combination of all four MRF-ASL parameters yielded the highest AUC (0.92±0.09). The other AUC values were: 0.77±0.17 for CBF_{1-compartment} alone, 0.78±0.17 for CBF_{2-compartment} alone, 0.75±0.14 for BAT alone, and 0.73±0.14 for T₁ alone. AUC value of ROC curve using the combination of MRF-ASL parameters for classification is considerably higher than any of the single parameter alone (p<0.001). No significant differences in AUC values were observed between the single-parameter results. It was also noted that different patients revealed different sensitivities in parametric maps.

CONCLUSIONS

Novel MRF-ASL technique allows simultaneous mapping of CBF, BAT, and tissue T₁ deficits in ischemic stroke. MRF-ASL derived parameters were predictive of NIHSS and mRS scores and capable of classifying lesion voxels with an accuracy of 92%.

Acknowledgements

No acknowledgement found.