Relaxation-normalized fast diffusion kurtosis imaging for semi-automatic segmentation of acute stroke lesion
Iris Yuwen Zhou1, Yingkun Guo1,2, Yu Wang3, Emiri Mandeville4, Suk-Tak Chan1, Mark Vangel1, Eng H Lo4, Xunming Ji3, and Phillip Zhe Sun1

1Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA, United States, 2Department of Radiology, Key Laboratory of Obstetric & Gynecologic and Pediatric Diseases and Birth Defects of Ministry of Education, West China Second University Hospital, Sichuan University, Chengdu, China, People's Republic of, 3Cerebrovascular Diseases Research Institute, Xuanwu Hospital of Capital Medical University, Beijing, China, People's Republic of, 4Neuroprotection Research Laboratory, Department of Radiology and Neurology, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA, United States


Kurtosis augments DWI for defining irreversible ischemic injury. However, long acquisition time of conventional DKI limits its use in the acute stroke setting. Moreover, the complexity of cerebral structure/composition makes kurtosis map heterogeneous, limiting the specificity of kurtosis hyperintensity to acute ischemia. With strongest correlation found between mean kurtosis and R1, we proposed the relaxation-normalized fast DKI approach to mitigate the kurtosis heterogeneity in normal brain with substantially reduced scan time. We further demonstrated that this approach enabled semi-automatic lesion segmentation and enhanced stratification of the heterogeneous DWI lesion, aiding the translation of fast DKI to the acute stroke setting.


Despite its widespread use, diffusion-weighted MRI (DWI) provides crude stratification of heterogeneous ischemic tissue injury1,2. Recent studies have shown that diffusion kurtosis imaging (DKI), measuring non-Gaussian diffusion, complements DWI for defining irreversible ischemic injury3,4. However, the conventional DKI acquisition time is relatively long, limiting its use in the acute stroke setting. In addition, the complexity of cerebral structure/composition makes kurtosis map heterogeneous, limiting the specificity of kurtosis hyperintensity to acute ischemia. Herein we developed relaxation-normalized fast DKI for the improved characterization of ischemic tissue injury, aiding the translation of fast DKI to the acute stroke setting.


Adult male Wistar rats were anesthetized throughout the experiments with 1.5-2.0% isoflurane. Multiparametric MRI was performed on two animal groups: normal rats (N=9) and stroke rats within 2 hrs after standard middle cerebral artery occlusion (MCAO, N=11) using a 4.7T Bruker scanner (Bruker Biospec, Billerica, MA). Multi-slice MRI (five 1.8-mm slices, FOV = 20x20 mm2, matrix = 48x48) was acquired with single-shot EPI. Fast DKI was acquired using three b-values: 0, 1000 (three directions), and 2500 (nine directions) s/mm2, gradient pulse duration/diffusion time (δ/Δ) = 6/20 ms, TR/TE = 2500/ 36.6 ms, 4 averages, scan time = 2 min 10 s5,6. T1-weighted images were acquired using an inversion recovery sequence, with seven inversion delays ranging from 250 ms to 3000 ms (TR/TE = 6500/14.8 ms). T2-weigthed SE images were obtained with two TEs of 30 and 100 ms (TR = 3250 ms). Images were analyzed in MATLAB (MathWorks, Natick, MA). We calculated mean diffusivity (MD) as described by Jensen et al.7.$$MD_{x,y,z}=\frac{(b_{1}+b_{3})D_{x,y,z}^{(12)}-(b_{1}+b_{2})D_{x,y,z}^{(13)}}{b_{3}-b_{2}}$$ where $$$D_{x,y,z}^{(ij)}=\frac{lnS(b_{i})/S(0)- lnS(b_{j})/S(0)}{b_{j}-b_{i}}$$$, i = 1, j = 2, 3, and b1=0, b2=1000, and b3=2500 s/mm2. We have $$$MD_{fast}=\frac{MD_{x}+MD_{y}+MD_{z}}{3}$$$. Mean kurtosis (MK) was obtained using the method described by Hansen et al.5 $$MK_{fast}=\frac{\frac{6}{15}(\sum_{i=1} ^3ln\frac{S(b_{3},\hat{n}^{(i)})}{S(0)}+2\sum_ {i=1}^3ln\frac{S(b_{3},\hat{n}^{(i+)})}{S(0)}+2\sum_ {i=1}^3ln\frac{S(b_{3},\hat{n}^{(i-)})}{S(0)})+6 \cdot b_{3} \cdot MD_{fast}}{b_3^2 MD_{fast}^{2}}$$ where $$$\hat{n}^{(1)}=(1,0,0)^{T}$$$, $$$\hat{n}^{(1+)}=(0,1,1)^{T}$$$ and $$$\hat{n}^{(1-)}=(0,1,-1)^{T}$$$, and similarly for i =2 and 3. Fractional anisotropy (FA) was estimated based on standard diffusion tensor model.

Results and Discussion

Fig. 1a shows parametric T1, T2, MD, FA, and MK maps from a representative normal Wistar rat brain. Fig. 1b compares the relationship between MK and MD, FA, relaxation rates of R1 and R2 in normal brain, and found strong correlation between MK and R1 (P<0.001). Based on this correlation, we used the univariate linear regression coefficients determined from R1 and MK (MKest= 1.36*R1-0.22) to generate relaxation-normalized mean kurtosis (RNMK=MK/MKest) maps. Fig. 2a shows that the RNMK approach could suppress intrinsic kurtosis heterogeneity in the intact brain. Importantly, it can also enhance ischemic kurtosis lesion segmentation over conventional MK map (Fig. 2b). Fig. 3 shows substantial MD and RNMK lesion mismatch in an acute stroke rat. We found significantly different RNMK and MD lesion volume, being 135±78 and 157±86 mm3, respectively (P<0.01, Paired-t test). Moreover, there was no significant difference in MD value from MD and RNMK lesions (0.62±0.04 µm2/ms vs. 0.61±0.03 µm2/ms, P>0.05, Paired-t test) while RNMK was significantly different between MD and RNMK lesions (1.52±0.15 vs. 1.70±0.13, P<0.01, Paired-t test).


Our study demonstrates that relaxation-normalized fast DKI reasonably corrects for intrinsic regional variation in cerebral kurtosis, enabling semi-automatic segmentation of ischemic kurtosis lesion with reduced scan time. The relaxation-normalized kurtosis analysis represents a promising approach to aid elucidation of the diagnostic value of DKI in stroke prior to its translation to the acute stroke setting.


The study was supported by grant from NIH/1R21NS085574.


[1] Sobesky J J. Cereb. Blood Flow Metab. 2012;32:1416-25.

[2] Yamada R, et al. Case Rep. Neurol. 2012;4:177-80.

[3] Jensen JH, et al. NMR Biomed. 2011;24:452-7.

[4] Cheung JS, et al. Stroke 2012;43:2252-4.

[5] Hansen B, et al. Magn. Reson. Med. 2013;69:1754-60.

[6] Sun PZ, et al. NMR Biomed. 2014;27:1413-8.

[7] Jensen JH, et al. NMR Biomed. 2010;23:698-710.


Fig. 1 (a) Comparison of multi-parametric R1, R2, mean diffusivity (MD), fractional anisotropy (FA) and mean kurtosis (MK) maps of a representative normal rat. (b) Linear regression analysis between kurtosis and multiple MRI indices of a representative normal rat.

Fig. 2 Comparison of R1, conventional MK map, MK map estimated from R1 map and relaxation-normalized MK (RNMK) map in (a) a representative normal rat and (b) a representative acute ischemic stroke rat.

Fig. 3 Ischemic tissue segmentation on the MD and RNMK maps. The size of ischemic lesions in MD or RNMK maps were semi-automatically defined using a threshold-based algorithm which counted the pixels with indices two standard deviations above the mean values of contralateral normal hemisphere.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)