Fast diffusion kurtosis imaging (DKI) with Inherent COrrelation-based Normalization (ICON) enhances automatic segmentation of heterogeneous diffusion MRI lesion in acute stroke

Abstract

Diffusion kurtosis imaging (DKI) has been shown to augment DWI for defining irreversible ischemic injury. However, the complexity of cerebral structure/composition makes kurtosis map heterogeneous, limiting the specificity of kurtosis hyperintensity to acute ischemia. We proposed an Inherent COrrelation-based Normalization (ICON) analysis to suppress the intrinsic kurtosis heterogeneity for improved characterization of heterogeneous ischemic tissue injury. Fast DKI and relaxation measurements were performed on normal (N=10) and stroke rats following MCAO (N=20). We evaluated the correlations between mean kurtosis (MK), mean diffusivity (MD), fractional anisotropy (FA) derived from fast DKI sequence and relaxation rates of R₁ and R₂, and found highly significant correlation between MK and R₁ (P<0.001). We showed that ICON analysis suppressed the intrinsic kurtosis heterogeneity in the normal cerebral tissue, enabling automated tissue segmentation in an animal stroke model. We found significantly different kurtosis and diffusivity lesion volumes, being 147±59 and 180±66 mm³, respectively (P=0.003, Paired-t test). The ratio of kurtosis to diffusivity lesion volume was 84±19% (P<0.001, One-sample t-test). We found relaxation normalized MK (RNMK) but not MD values significantly different between kurtosis and diffusivity lesions (P<0.001, ANOVA). Our study showed that fast DKI with ICON analysis provides a promising means for demarcating heterogeneous DWI stroke lesion.

Graphical Abstract

Kurtosis augments DWI for defining irreversible ischemic injury. However, the complexity of cerebral structure/composition makes kurtosis map heterogeneous, limiting the specificity of kurtosis hyperintensity to acute ischemia. With strong correlation found between mean kurtosis and R₁, we proposed an Inherent COrrelation-based Normalization (ICON) approach to mitigate the kurtosis heterogeneity in normal brain with substantially reduced scan time. We further demonstrated that this approach enabled automatic lesion segmentation and enhanced stratification of heterogeneous DWI lesion, aiding the translation of fast DKI to the acute stroke setting.

1. INTRODUCTION

Diffusion-weighted imaging (DWI) is sensitive to early ischemic tissue injury, and can detect ischemic lesion within minutes of the ischemic insult. Given its advantage in early detection of acute stroke, DWI has become one of the most widely used MRI techniques for stroke imaging (1–4). Various mismatch paradigms, including perfusion/diffusion, angiogram/diffusion and clinical/diffusion where the NIH Stroke Scale (NIHSS) used as a clinical indicator of infarction volume, have been proposed to guide stroke treatment beyond the standard therapeutic window (5–8). However, it has been recognized that ischemic tissue damage within the DWI lesion is heterogeneous (9–15). A portion of the DWI lesion represents injury that is reversible when reperfusion is prompt and effective, and the extent of DWI lesion reversibility is independently associated with improved outcome (16–19). As such, the underlying assumption that the DWI lesion identifies the irreversibly injured ischemic core is oversimplified, and advanced imaging techniques are urgently needed to refine diffusion stroke MRI (20).

Diffusion kurtosis imaging (DKI) has been introduced as a method to quantify non-Gaussian diffusion and has been applied to a number of neurological disorders including Parkinson’s disease, Huntington’s disease, traumatic brain injury and stroke (21–31). It has been shown that there is significant mismatch between mean diffusivity (MD) and mean kurtosis (MK) lesions in an animal model of acute stroke (32). In addition, diffusivity lesion without kurtosis abnormality responds favorably to early reperfusion, where little change is found in the kurtosis lesion, suggesting that the kurtosis abnormality captures the more severely damaged ischemic core and hence, stratifies the conventional DWI lesion. However, unlike the relatively homogeneous trace diffusion image, the complexity of cerebral structure and composition leads to a heterogeneous kurtosis map, in which the specificity of kurtosis abnormality to ischemia is compromised, making automatic lesion segmentation somewhat difficult. A means to minimize the intrinsic cerebral tissue kurtosis variation would thus enhance the conspicuity of the ischemic kurtosis lesion.

The conventional DKI protocol requires collecting DWI images with multiple b-values along at least fifteen directions, resulting in relatively long acquisition time (21). Recently, a fast DKI protocol has been introduced by Hansen et al. for the rapid mapping of MD and MK (33,34). The diffusivity and kurtosis maps obtained with the fast DKI agreed well with those estimated from conventional DKI using the diffusion tensor model (MD_tensor and MK_tensor) in normal brain (33), in glioma (35) and importantly in acute stroke brain (33,34,36). We have further shown that kurtosis is one of the most sensitive indices in capturing acute ischemic lesion (37). Herein we developed a relaxation-normalized fast DKI approach to refine acute ischemic lesion stratification. Briefly, we evaluated the correlation between cerebral MK and multiple MRI indices including MD, fractional anisotropy (FA), and relaxation (R₁, R₂). There were significant correlation between MK and R₁, MK and R₂, and MK and FA, with R₁ showing the highest correlation coefficient. Because MD and FA may alter substantially during acute stroke while the change in relaxation rate is relatively small, we proposed an Inherent COrrelation-based Normalization (ICON) analysis and validated it in suppressing the intrinsic kurtosis heterogeneity. We demonstrated that the ICON approach enabled automatic ischemic lesion segmentation during acute ischemic stroke, superior to the MK-based lesion analysis. In summary, our study developed relaxation-normalized fast kurtosis imaging for characterizing heterogeneous ischemic tissue injury, aiding the translation of fast diffusion kurtosis imaging to the acute stroke setting.

2. MATERIALS AND METHODS

Animals

All experiments have been approved by the Institutional Animal Care and Use Committee. Adult male Wistar rats (Charles River Laboratory, Wilmington, MA) were anesthetized throughout the surgery and the MRI experiments with 1.5–2.0% isoflurane/air mixture. Animal body temperature was maintained by a circulating warm water jacket positioned around the torso. We imaged ten normal rats (N=10) and twenty three stroke rats following the standard permanent middle cerebral artery occlusion (MCAO, N=23). Briefly, MCAO was induced by inserting a 4–0 silicon-coated nylon suture (Doccol Corp, Sharon, MA) into the lumen of the right internal carotid artery, and then advancing it to block the origin of the middle cerebral artery. Stroke rats were then transferred to the MRI scanner after the surgery and imaged 1.5 to 2 hours after the onset of occlusion. Peripheral oxygen saturation (SpO2) and heart rate were continuously monitored (Nonin Pulse Oximeter 8600, Plymouth, MN). The SpO2 was not significantly different between the two groups, being 95 ± 4 % (normal) vs. 92 ± 3 % (MCAO). Three animals were excluded from analysis due to failed MCAO surgery.

MRI

MRI data were acquired using a 4.7 Tesla small-bore MRI scanner (Bruker Biospec, Billerica, MA). We used a dual RF coil setup, including a 70 mm volume transmitter coil and an actively-decoupled 20 mm surface receiver coil. During MRI, the animals were placed on a plastic cradle with the head fixed with a tooth bar and ear bars to minimize motion. Multi-slice MRI (5 slices, slice thickness/gap = 1.8/0.2 mm, field of view = 20×20 mm², acquisition matrix = 48×48) was acquired with single-shot echo-planar imaging (EPI) (receiver bandwidth = 225 kHz), with the central slice positioned at 2 mm posterior to bregma. Fast DKI was acquired using three b-values: one b = 0 s/mm² image, nine b = 1000 s/mm² and nine b = 2500 s/mm² images along directions of n̂⁽ⁱ⁾, n̂⁽ⁱ⁺⁾, n̂⁽ⁱ⁾ (i.e. n̂⁽ⁱ⁾ = (0,1,1)^T, n̂⁽¹⁺⁾ = (0,1,1)^T, and n̂⁽¹⁻⁾ = (0,1,−1)^T, and similarly for i =2 and 3) (38). Note that the superscript i in n̂⁽ⁱ⁾ labels the position of the “1” while in n̂⁽ⁱ⁺⁾ and n̂⁽ⁱ⁻⁾ it labels the position of the “0” (33,34). The other parameters of the fast DKI approach were gradient pulse duration/diffusion time (δ/Δ) = 6/20 ms, repetition time (TR)/echo time (TE) = 2500/36.6 ms, 4 averages, scan time = 3 min 10 s. In addition, T₁-weighted images were acquired using an inversion recovery sequence, with seven inversion delays ranging from 250 to 3000 ms (TR/TE = 6500/14.8 ms, 4 averages, scan time = 3 min 38 s); T₂-weigthed SE images were obtained with two TEs of 30 and 100 ms (TR = 3250 ms, 4 averages, scan time = 26 s). Perfusion imaging was acquired with amplitude modulated arterial spin labeling MRI (TR/TS/TE = 6500/3250/14.8 ms, 32 averages, scan time = 7 min).

Image and Data Analysis

Images were analyzed in MATLAB (MathWorks, Natick, MA). MD and MK maps were obtained by linear combination of log diffusion signals, as described previously (33,34,36,37,39). FA was estimated based on standard diffusion tensor model. In addition, parametric T₁ map was derived using least-squares fitting of the signal as functions of inversion time (I = I₀ ⌊1 −(1−η)e^−TI/T₁⌋)), where η is the inversion efficiency, TI is the inversion time, and I₀ is equilibrium signal. T₂ map was calculated as ${\mathrm{T}}_2=\frac{{\text{TE}}_2-{\text{TE}}_1}{ln(\mathrm{I}({\text{TE}}_1)/\mathrm{I}({\text{TE}}_2))}$, where I(TE_1,2) are T₂-weighted signals obtained at different echo times (TE= 30 and 100 ms). The relationship between MK and other MRI indices (T₁, T₂, MD and FA) was evaluated using Pearson’s correlation, excluding ventricular area with MD value above 1 μm²/ms. An estimated MK map (MK_est) was calculated using the regression coefficients determined from the univariate linear regression analysis between R₁ and MK per pixel (i.e. MK_est= C₀+C₁*R₁, where C₀ and C₁ are regression coefficients) and then the relaxation-normalized MK (RNMK) map was calculated as RNMK=MK/MK_est. Cerebral blood flow (CBF) map was obtained from amplitude modulated arterial spin labeling MRI (40) and derived as $\text{CBF}=\frac{\mathrm{\lambda }\left({\mathrm{I}}_{\text{ref}}-{\mathrm{I}}_{\text{tag}}\right)}{2\mathrm{\alpha }·{\mathrm{I}}_{\text{ref}}}·\frac{{\mathrm{e}}^{\mathrm{w}/{\mathrm{T}}_{1\mathrm{a}}}}{{\mathrm{T}}_{1\mathrm{w}}}$, where I_tag is the label image, I_ref is the reference image, λ is the brain-blood partition coefficient for water, α is the degree of inversion with transient time correction, w is the post-labeling delay, and T_1a is the arterial blood longitudinal relaxation time (41).

Results were expressed as mean ± standard deviation (S.D.). We evaluated the association between MK and relaxation, MD and FA with univariate and multiple linear regressions, and P-values of less than 0.05 were considered statistically significant. The size of ischemic lesions in MD, MK or RNMK maps were automatically defined using a K-means clustering-based algorithm (42). Briefly, a 1D vector of all the pixels within the brain was formed and clustering algorithm used an iterative refinement technique to partition the data points in the vector into two clusters (lesion vs. normal tissue) based on their intensities.

3. RESULTS

Fig. 1 shows parametric brain T₁, T₂, MD, FA, and MK maps from a representative normal Wistar rat. Because cerebral spinal fluid (CSF) has significantly higher water content, the ventricle region appears hyperintense in T₁, T₂ and MD images. In addition, CSF displays hypointensity in FA and MK maps due to little anisotropic/restricted diffusion. Whereas the contrast between striatum or corpus callosum regions and cortical region was relatively small in T₂ and MD maps, it was distinct in T₁, FA and MK maps. This suggests that kurtosis, a measure of non-Gaussian diffusion, is partially associated with diffusion fractional anisotropy, a widely-used measure of gross white matter microstructure (43–45) along with longitudinal relaxation rate, which has been shown to be associated with semisolid macromolecules and water content (46,47).

Fig 1.

Comparison of T₁, T₂, mean diffusivity (MD), fractional anisotropy (FA) and mean kurtosis (MK) maps of a representative normal rat.

We evaluated the relationship between MK and other MRI indices using Pearson’s correlation with a Student’s t-distribution, excluding ventricular regions using an MD threshold of 1 μm²/ms (Fig. 2a). There was a weak correlation between MK and MD (R²=0.11, P<0.001), consistent with previous report suggesting that diffusivity and kurtosis are two distinct diffusion-based indices for characterizing tissue microstructure (48). The pixel-wise analysis also showed a weak correlation between MK and FA (R²=0.17, P<0.001), moderate correlation between MK and R₂ (=1/T₂, R²=0.37, P<0.001) and strong correlation between MK and R₁ (=1/T₁, R²=0.70, P<0.001). The correlation between MK and R₁ was significantly higher than that between MK and MD, FA or R₂ in normal animals (N=10, ANOVA with multiple comparison test, P<0.001). Because MD and FA may change substantially during acute stroke, we compared the univariate linear regression of MK and R₁ versus multiple linear regression of MK with R₁ and R₂ in all the normal animals (Fig. 2b). We found that the coefficient of determination (i.e., R²) was 0.62 ± 0.08 (MK and R₁, P<0.001) and 0.62 ± 0.08 (MK, R₁ and R₂, P<0.001), showing little difference (Two-sample t-test, P=0.92). Hence, we chose the univariate linear regression of MK and R₁ to correct for intrinsic kurtosis heterogeneity in intact cerebral tissue.

Fig 2.

a) Univariate linear regression analysis between mean kurtosis and multiple MRI indices, including MD, FA, R₂ (=1/T₂) and R₁ (=1/T₁) of a representative normal rat. b) Multiple linear regression analysis between mean kurtosis and R₁, R₂ of the same animal.

Fig. 3 compares the routine MK and relaxation-normalized MK maps in a normal Wistar rat. Both the relaxation R₁ (i.e., 1/T₁) map (Fig. 3a) and the MK map (Fig. 3b) show hyperintensity in striatum and corpus callosum. Fig. 3c shows the estimated MK map (MK_est) using the regression coefficients determined from univariate linear regression analysis between R₁ (Fig. 3a) and MK (Fig. 3b) per pixel (MK_est= 1.42*R₁-0.26). Fig. 3d shows the proposed relaxation-normalized MK map (i.e. RNMK=MK/MK_est), which was substantially more homogeneous than the MK map without relaxation normalization (Fig. 3a). For all the normal animals studied, the coefficient of variation (COV, i.e., S.D./mean) was 23.6 ± 2.3% and 12.1 ± 2.8% for MK and RNMK maps, respectively. This represents 48.6 ± 10.2% decrease in COV, showing that the ICON analysis removes a substantial portion of kurtosis heterogeneity which is intrinsic to normal cerebral tissue.

Fig 3.

Fast DKI with Inherent COrrelation-based Normalization (ICON) analysis in a representative normal rat. a) R₁ map. b) MK map with substantial regional variation. c) MK map estimated from R₁ map. d) Relaxation-normalized MK (RNMK) map (RNMK=MK/MK_est), which was substantially more homogeneous.

We further evaluated the ICON analysis in a representative acute stroke animal within 2 hrs after ischemia. R₁ (Fig. 4a) and R₂ (Fig. 4b) maps show relatively small change during acute ischemic stroke, with a reduction of 1.8 % in R₁ and 0.1% in R₂ in the ischemic lesion. In comparison, substantial reduction in MD map was observed in the ischemic lesion (Fig. 4c). The diffusivity lesion was 42.7 mm³ for the slice. The MK map is heterogeneous, particularly in the internal capsule and striatum, which leads to overestimated kurtosis lesion (Fig. 4d). While the ipsilateral cortex shows relatively increased kurtosis from that of the contralateral cortex, the K-means clustering algorithm missed part of the cortex lesion but incorrectly included the contralateral internal capsule and ipsilateral striatum to the kurtosis lesion (Fig. 4d, arrows), resulting in kurtosis lesion substantially larger than the diffusivity lesion (54.2 vs. 42.7mm³). Fig. 4e demonstrates significant correlation between R₁ and MK indices from the contralateral normal brain (R²=0.59, P<0.001). For the contralateral brain, MK and RNMK values were found to be 0.67 ± 0.13 and 1.00 ± 0.12, respectively. This represents a relative COV decrease of 42%, similar to that found in intact brains. The kurtosis lesion was segmented from RNMK map, using the K-means clustering algorithm (Fig. 4f). Compared to the lesion segmented in Fig. 4d which inaccurately includes part of the contralateral normal tissue, the ICON analysis corrected for such error and resulted in a kurtosis lesion size of 41 mm³, which is close to but slightly smaller than the MD lesion (as seen in Fig. 4c and 4f).

Fig 4.

Relaxation-normalized fast DKI of a representative rat after acute ischemic stroke. a) R₁ map. b) R₂ map. c) MD map with outlined ischemic lesion. d) MK map with outlined ischemic lesion. e) Linear regression analysis of R₁ and MK in the contralateral normal brain. f) RNMK map with outlined ischemic lesion.

We further tested relaxation-normalized kurtosis analysis for multi-slice segmentation in acute stroke rats. Fig. 5a shows the outlined diffusivity and kurtosis lesions superimposed on a scout scan, in which diffusivity and kurtosis lesions appear in red and green, respectively. Fig. 5b shows substantial kurtosis/diffusivity mismatch. Fig. 5c compares diffusivity and kurtosis lesion volumes in all stroke animals. Specifically, we found kurtosis lesion significantly smaller than that of diffusivity lesion (147±59 vs. 180±66 mm³, P=0.003, Paired-t test). The ratio of kurtosis to diffusivity lesion volumes was 84±19% (P<0.001, One-sample t-test). Moreover, we compared the CBF, MD, RNMK values in different brain regions including contralateral normal, diffusivity lesion, kurtosis lesion and their mismatch (Table 1). MD was 0.63±0.06 and 0.65±0.06 μm²/ms in diffusivity and kurtosis lesions (n.s., ANOVA) while their RNMK values were significantly different (1.60±0.27 vs. 1.77±0.21, P<0.05, ANOVA). The kurtosis/diffusivity lesion mismatch had a relatively higher MD value (p<0.05, ANOVA) while its RNMK value was significantly lower than that of diffusivity lesion (p<0.001, ANOVA) and kurtosis lesion (p<0.001, ANOVA). Our results demonstrated that relaxation-normalized kurtosis imaging effectively reduced the intrinsic kurtosis heterogeneity, enabling automatic tissue segmentation of acute stroke DKI.

Fig 5.

Ischemic tissue segmentation using the proposed relaxation-normalized kurtosis MRI. a) MD lesion (red) vs. RNMK lesion (green). b) 3-D overlay of diffusion and kurtosis lesions. c) Comparison of ischemic diffusivity and kurtosis lesions.

Table 1.

Comparisons of cerebral blood flow (CBF), MD and RNMK values in different brain regions.

N=20	Contralateral Normal Region	Diffusivity Lesion	Kurtosis Lesion	Kurtosis/Diffusivity Lesion Mismatch
CBF (ml/g/min)	1.27 ± 0.42	0.64 ± 0.40***	0.59 ± 0.37***	0.79 ± 0.51**
MD (μm2/ms)	0.86 ± 0.03	0.63 ± 0.06***	0.65 ±0.06***	0.68 ± 0.06***†
RNMK	1.00 ± 0.01	1.60 ± 0.27***	1.77 ± 0.21***§	1.17 ± 0.15*§§§†††

One-way ANOVA with Bonferroni’s multiple comparisons test was performed.

^***p<0.001,

^**p<0.01,

^*p<0.05 significantly different from contralateral normal region;

^§§§p<0.001,

^§p<0.05 significantly different from diffusivity lesion;

^†††p<0.001,

^†p<0.05 significantly different from kurtosis lesion. Note all the measurements passed D’Agostino-Pearson normality test, indicating Gaussian distribution.

4. DISCUSSION

Our results demonstrated significant correlation between mean kurtosis and longitudinal relaxation, and that relaxation-normalized fast DKI reduces the intrinsic kurtosis heterogeneity, facilitating automatic segmentation of heterogeneous ischemic tissue. In addition, we showed that there was no statistically significant difference in mean diffusivity between diffusivity and kurtosis lesions, consistent with prior studies that showed the DWI lesion cannot be reliably characterized based solely on the severity of diffusion coefficient deficit (16). Our prior work showed that in an animal model of acute stroke, diffusivity and kurtosis lesion mismatch responds favorably to early reperfusion with little change in the kurtosis lesion, demonstrating that kurtosis index may resolve heterogeneous DWI lesion (32). However, because of the intrinsic heterogeneity between structures in the MK map, lesions were manually outlined. Our study here developed a relaxation-normalized MK analysis for automatic lesion segmentation, and confirmed significant acute stroke diffusivity and kurtosis lesion mismatch. It is important to note that kurtosis change is sensitive to complex origins including multi-compartment diffusion, restricted diffusion and tortuosity, etc. It is likely that the diffusivity/kurtosis lesion mismatch is subject to intracellular/extracellular net water shift following trans-membrane ion channel dysfunction while kurtosis lesion undergoes pronounced intracellular tortuosity and viscosity elevation, precursors of the irreversible cell damage (49–51).

The proposed ICON analysis utilized the strong correlation between MK and R₁ to suppress the intrinsic kurtosis heterogeneity. We evaluated the R₁ and R₂ rates of all the stroke animals and found that ischemic lesion as defined by MD deficit showed relatively small R₁ decrease compared to that of the normal side, being only 3.6 ± 2.2 % (P<0.01), and no detectable changes in R₂ (P=0.74), consistent with prior studies (52,53). Meanwhile, ICON analysis shows a substantial increase in the lesion with an average change of around 17% in the kurtosis/diffusivity mismatch area and it goes to as high as 77% in the kurtosis lesion and 60% in the diffusivity lesion. As such, the ICON approach performs well for acute stroke imaging.

We used a fast DKI protocol to minimize the acute stroke scan time (33,34). Our previous work has demonstrated significant correlation of kurtosis measured by the routine DKI protocol and the fast DKI method in an animal model of acute stroke (36). Recently we also showed that MK is one of the most sensitive parameters in revealing acute ischemic injury when compared to radial and axial kurtosis, MD and FA (37). Although our study acquired additional six images at b=1000 s/mm² for fractional anisotropy calculation, these additional images can be omitted in the future due to the weak correlation between MK and FA.

Our study demonstrated significant yet relatively weak correlations between kurtosis, FA and transverse relaxation rate. More importantly, we showed that longitudinal relaxation alone accounts for approximately 49% of kurtosis heterogeneity in normal brain. Kurtosis measures non-Gaussian diffusion behavior, which is strongly related to tissue microstructure (e.g., macromolecules, membranes and fibers) that creates diffusion barriers and compartments (25,39). A large diffusional kurtosis suggests a high degree of diffusional heterogeneity and/or microstructural complexity. This is particularly evident in brain, where water diffusion is strongly restricted by myelinated axons. On the other hand, longitudinal relaxation time T₁ highly correlates with magnetization transfer ratio (MTR), an index sensitive to macromolecules (47). T₁ is also closely related to tissue myelination (54,55) and in particular, the cholesterol that is bound to myelin (56). In addition, T₁ was found to be associated with tissue water content (46,57). It is likely that macromolecules, myelin and/or water content may affect multiple MRI indices of healthy brain, which may partially explain the correlation among the measurements. Additional MRI indices such as magnetization transfer (MT) and amide proton transfer MRI which contain information about the chemical and physical environments of macromolecular protons in tissue may further enhance the correlation (58–60) and hence improve suppression of intrinsic heterogeneity in kurtosis map and specificity of MK changes to ischemic tissue injury. It is necessary to point out that multi-parametric MRI acquisition beyond the highly correlated T₁ and kurtosis MRI will prolong the scan time, limiting its use during acute stroke imaging. Techniques such as fingerprinting MRI may make the proposed approach more applicable in the emergency settings (61).

Accurate identification of the salvageable ischemic tissue is crucial in the clinical settings, particularly with the recent advancement in embolectomy. In this work, we studied acute ischemic tissue injury within 2 hours from occlusion using a well-established filament MCAO model. This time window is similar to the hyperacute phase of stroke patients, which is the most critical period for effective treatment. Although reasonably reproducible, the filament model induces severe hypoperfusion and, arguably, significantly reduces the salvageable tissue and shortens the treatment window. This limitation may be reasonably addressed by using an embolic stroke model, albeit its moderately higher variability (62). Another advantage of using the embolic model is that it enables evaluation of thrombolytic therapy that mimics recanalization in stroke patients. More importantly, our method can be used to study the irreversible and reversible ischemic lesion in the embolic model with thrombolytic treatment, especially beyond the hyperacute window before significant T₁ change occurs, potentially extending the treatment window. On the other hand, there may be substantial relaxation changes in the ischemic lesion at extended stroke onset time, which can be largely attributable to edema formation. The performance of ICON approach in subacute stroke imaging needs to be investigated in the future. In the current study, no histology was obtained because commonly used Hematoxylin and eosin (H&E) and TTC staining techniques cannot reliably identify acute ischemic tissue injury. Notably, before we can investigate histological changes, we need to develop a reliable lesion segmentation approach to guide histology, which is the goal of our current study. It would be interesting to confirm the tissue classification and reversibility of ischemic brain using more sensitive immunohistological techniques in the future.

In this study, we chose a relatively short diffusion time, which is desirable given that brain T₂ at 4.7 T is shorter than that at 1.5 and 3 T. Recently, Baron et al. found reduced diffusion contrast in stroke patients under short diffusion times (63). Latt et al. showed that DWI acquired with long diffusion time is susceptible to intra- and extra-cellular water exchange, especially in regions with compromised cell membranes (64). In addition, tissue with intact but functionally impaired and nearly impermeable membranes shows significantly reduced ADC. Consequently, long diffusion time may enhance diffusion contrast with elevated heterogeneity, for which the proposed ICON analysis might provide crucial information for tissue characterization. To summarize, the proposed relaxation-normalized fast kurtosis imaging is promising to aid the evaluation of fast DKI and its translation to the acute stroke setting.

5. CONCLUSION

Our study demonstrated that fast kurtosis imaging with Inherent COrrelation-based Normalization (ICON) reasonably corrects for the intrinsic regional kurtosis variation, enabling automatic segmentation of ischemic kurtosis lesion. We confirmed significant kurtosis/diffusivity lesion mismatch, which is a promising approach to elucidate the diagnostic value of fast DKI in acute stroke prior to its translation to the clinical setting.

ABBREVIATIONS

ADC

Apparent Diffusion Coefficient

CBF

Cerebral Blood Flow

COV

Coefficient of Variation

CSF

Cerebrospinal Fluid

DKI

Diffusion Kurtosis Imaging

DWI

Diffusion-weighted Imaging

EPI

Echo Planar Imaging

FA

Fractional Anisotropy

MCAO

Middle Cerebral Artery Occlusion

MD

Mean Diffusivity

MK

Mean Kurtosis

TE

Echo Time

TR

Repetition Time