Diffusion kurtosis imaging for characterizing tumor heterogeneity in an intracranial rat glioblastoma model

The utility of diffusion kurtosis imaging (DKI) for assessing intra‐tumor heterogeneity was evaluated in a rat model of glioblastoma multiforme. Longitudinal MRI including T2‐weighted and diffusion‐weighted MRI (DWI) was performed on six female Fischer rats 8, 11 and 14 days after intracranial transplantation of F98 cells. T2‐weighted images were used to measure the tumor volumes and DWI images were used to compute diffusion tensor imaging (DTI) and DWI based parametric maps including mean diffusivity (MD), mean kurtosis (MK), axial diffusivity (AD), axial kurtosis, radial diffusivity, radial kurtosis, fractional anisotropy (FA) and kurtosis fractional anisotropy (KFA). Median values from the segmented normal contralateral cortex, tumor and edema from the diffusion parameters were compared at the three imaging time points to assess any changes in tumor heterogeneity over time. ex vivo DKI was also performed in a representative sample and compared with histology. Significant differences were observed between normal cortex, tumor and edema in both the DTI and DKI parameters. Notably, at the earliest time point MK and KFA were significantly different between normal cortex and tumor in comparison with MD or FA. Although a decreasing trend in MD, AD and FA values of the tumor were observed as the tumor grew, no significant changes in any of the DTI or DKI parameters were observed longitudinally. While DKI was equally sensitive to DTI in differentiating tumor from edema and normal brain, it was unable to detect longitudinal increases in intra‐tumoral heterogeneity in the F98 model of glioblastoma multiforme.

GBM in rats, 2 including a high degree of heterogeneity, invasiveness and diffused boundaries. 3 It has been used to assess chemo 4,5 and radiation 6 therapy and has also been used in MR studies including spectroscopy, 7 diffusion 8-10 and perfusion MRI. 6,11 Diffusion-weighted magnetic resonance imaging (DWI) has been widely used to quantify the random motion of water molecules in biological tissues. [12][13][14] Standard analytical models processing diffusion-weighted MRI data for computation of the apparent diffusion coefficient (ADC) values assume water displacement in the tissue (voxel of interest) follows a Gaussian statistical distribution, similar to the water diffusion observed in homogeneous liquids. However, it is well known that the assumption of Gaussian distribution fails in in vivo conditions due to the inherent heterogeneity from the presence of various tissue compartments, including different cell types, cell morphologies, extracellular matrix and blood. 15 Diffusion kurtosis imaging (DKI) is a dimensionless metric that quantifies how much the water diffusion deviates from a Gaussian distribution due to cellular membranes, intra-and extracellular compartments and tissue structure. [15][16][17] Thus, the diffusion of water molecules in homogeneous liquids will follow a Gaussian distribution with a kurtosis of zero. In tissues where diffusion is mostly hindered and restricted, water molecules will more likely diffuse short distances around the initial position in a time t, leading to a sharper statistical distribution and a positive kurtosis.
DKI has been used to assess white matter (WM) damage and myelin density. 17 Preclinical DKI studies include infarct, 18 traumatic brain injury 19 and Alzheimer's disease, 20 type 2 diabetic ischemic stroke 21 and acute alcohol intoxication. 22 DKI has also been reported to aid in assessing microstructural heterogeneity in tumors and its degree of diffusion restriction. It has been used in grading of human gliomas whereby higher mean kurtosis (MK) and lower mean diffusivity (MD) values were noted in high-grade solid tumors with increased cellularity. 23 Increased cellularity and the presence of spindle-shaped cells led to a higher kurtosis and lower diffusivity in colorectal tumor xenografts. 24 Although promising, none of the published studies have assessed longitudinal changes in kurtosis parameters of the tumor for assessing changes in tumor tissue heterogeneity with regard to the microenvironment and cellular components as the tumor grows. Therefore, we performed a longitudinal study in a rat F98 brain tumor model to assess whether changes in DKI parameters can better assess tumor heterogeneity as the tumor volume increases over time. Scientific, Waltham, MA, USA). The cells were maintained at 37 C in a 5% CO₂ humidified atmosphere. Cells were passaged twice-weekly at 1 x 10 5 per T-75 flask and terminated after the fifth passage to avoid any chance of further mutations. Cells were tested bi-monthly for mycoplasma.

| Brain tumor model
In vivo studies on rats were conducted in compliance with the UK Home Office (Animals Scientific Procedures Act 1986) and with the ethical approval of the local committee of the University of Liverpool. Six F344 female (100-120 g) Fischer rats (Charles River, Margate, UK) were injected with 50 000 F98 cells suspended in 5 μL serum-free DMEM culture medium. The injection was performed in an aseptic environment using sterile tools. The rat was maintained under surgical anesthesia using a 3% isoflurane in O 2 gas mixture. Rats were given subcutaneous injections of antibiotics (5 mg/kg, 25 mg/mL enrofloxacin, 2.5% Baytril, Bayer, Leverkusen, Germany) and analgesia (0.3 mg/mL buprenophine, Vetergesic, Ceva Animal Health, Amersham, UK) before the surgery, and 2 mL saline after the surgery. The rat was maintained in a three-point stereotaxic frame, the head was shaved and a small incision allowed access to the skull. A burr hole was drilled through the skull 3 mm right and 3 mm posterior from the bregma and the cells were injected 2.5 mm deep into the cerebral cortex. After the surgery, the skin was sutured, and the animal was returned to its cage for recovery. Three animals were housed together in a cage with stimulation objects and free access to food and water, which was provided ad libitum and the animals were kept in a 12-hour day/light cycle.

| MRI acquisition
MRI scans were performed at 9.4 T on a Bruker Biospec (Bruker BioSpin, Ettlingen, Germany). Signal was generated using an 86 mm transmission birdcage coil and detected by a four-channel phased array surface coil. The rats were anesthetized with 2% isoflurane in O 2 and the respiration rate and body temperature were monitored using an abdominal motion sensor and a rectal probe (SA Instruments, Stony Brook, NY, USA). The body temperature was maintained at 35 C by a hot water blanket and the respiration rate at 50-60 inspirations per minute. Each MRI experiment consisted of a localizer scan, followed by an anatomical T 2 -weighted sequence and a DWI sequence.
In vivo MR images were acquired longitudinally on days 8, 11 and 14 after inoculation of tumor cells to assess changes in the tumor microenvironment with DKI using a minimum of three time points (the early, mid and late tumor stages). These time points were also chosen to comply with UK Home Office guidelines of not subjecting the animal to undue stress of multiple anesthesia sessions or exceeding the severity limits on animal health. A multi-slice T 2 -weighted sequence was acquired to locate the tumor using a fast spin echo sequence with the following parameters: TE/TR = 33/5000 ms, RARE factor = 8, matrix = 256 x 256, FOV = 40 x 20 mm, 38 slices, scan duration = 2 minutes 38 seconds. DKI was performed using a respiratory-gated EPI-DTI sequence with the parameters: TE/TR = 23/2500 ms, five averages, four EPI segments, matrix = 128 x 64, FOV = 40 x 20 mm, 38 slices, voxel resolution = 0.3 x 0.3 x 0.3 mm 3 , δ/Δ = 4/11 ms, 15 directions, b-values = 0-1000-2000 s/mm 2 , three b0 images, 27.5 minutes. The total scan duration for each experiment was 60 minutes. Animals were rehydrated with 1 mL saline injected subcutaneously after each MRI session.

| Image processing and statistical analysis
The brain was manually segmented on the b = 0 s/mm 2 images from the diffusion-weighted datasets using ITK-SNAP (www.itksnap.org).
Tumors were manually segmented on the T 2 -weighted images to assess tumoral growth. Tumor growth rate was calculated from the logarithm of the volume ratio from day 8 to day 14, and volumetric doubling time was then calculated using the exponential growth model. 25 Diffusion and kurtosis parametric maps were calculated using the Diffusional Kurtosis Estimator (DKE) software (Medical University of South Carolina, SC, USA). A characteristic T 2 -weighted image of a typical rat 11 days after tumor cells injection and its corresponding parametric maps are shown in Figure 1. No corrections were made for geometric distortions or eddy current effects. Volumes of interest (VOIs) corresponding to the whole tumor, the whole peritumoral edema and the contralateral normal-appearing healthy brain parenchyma cortex were also segmented using ITK-SNAP and the binary masks were overlaid on the parametric maps. The contralateral normal brain healthy cortex VOI was segmented by selecting a region of left frontal cortex for every slice containing the tumor. As the contralateral normal brain microstructure is unlikely to change due to the presence of the tumor, it was used as a reference with the hypothesis that no significant changes in normal brain will be observed, while changes in tumor heterogeneity will lead to changes in DKI parameters. Care was taken to keep normal brain F I G U R E 1 A, representative T 2 -weighted image and B, its corresponding diffusivity and C, kurtosis parametric maps. The red arrow indicates the tumor location on the T 2 -weighted image VOI as big and as close as possible to the first imaging time point in each animal, and during longitudinal studies. Typical VOIs are shown in Figure 2B. Histograms of the parameter value distribution in the tumoral, edematous and cortical regions (Figure 2A,C) were generated using MATLAB (Mathworks, Natick, MA, USA). Mean, median and standard deviation values were calculated for each parameter using Origin (OriginLab, Northampton, MA, USA). A Wilcoxon signed-rank test was used to compare the tumor diffusion and kurtosis median values with the peritumoral edema and contralateral cortex. A Friedman test was used to compare the longitudinal data. A P-value of .05 or less was considered to be significantly different between the groups.

| Tissue collection
Animals were euthanized 1 day after the last MRI session using an overdose of 3 mL/kg pentobarbital sodium (Euthatal, Merial Animal Health, Harlow, UK) injected intraperitoneally. An incision was performed along the mid-ventral line through the abdomen to severe the aorta under the diaphragm. A midline thoracotomy gave access to the heart. A 25-gauge needle connected to an extension tube was clamped to the left ventricle of the heart to perfuse with 50 mL saline followed by 75 mL 4% Formalin (Sigma-Aldrich). Following fixation, brains were collected and suspended in 4% Formalin.

| Ex vivo MRI
Ex vivo MR images of the brain suspended in perfluoropolyether oil (Fomblin, Solvay, Brussels, Belgium) were acquired using the same T 2weighted coronal fast spin echo sequence as the in vivo protocol except that 25 averages were used (scan duration = 1 hour 12 minutes). DWI was carried out using the same EPI-DTI sequence that was used in vivo with 25 averages (scan duration = 3 hours 26 minutes).

| Histology
The brain sample that was used for ex vivo DKI study was embedded in paraffin until sectioning after the DKI study. Hematoxylin and eosin (H&E) staining was performed on 4 μm coronal sections across the tumor. The sections closely matching the ex vivo imaging slice were qualitatively analyzed and the extent of cell density and cellular organization was based on visual assessment of staining.  Table 1 lists the mean and standard deviation values of the diffusivity and kurtosis parameters in the six rats.

| RESULTS
Tumors were observed on the MD maps with a concentric hyperintense structure composed of the peritumoral edema and the necrotic core ( Figure 2B). The tumor appears hyperintense on the MK maps ( Figure 2D). The axial (AD) and radial diffusivity (RD) maps showed a concentric structure similar to that observed on the MD maps formed of high diffusivity in the peritumoral edema and necrotic core. Likewise, the axial

| Tumor vs. contralateral cortex
A Wilcoxon signed-rank test showed a significantly higher MD in the tumor compared with the contralateral cortex from day 11 (Z = 2.097, P = .036) ( Figure 4A). Similar to MD, the RD was significantly higher in the tumor than the contralateral cortex, as illustrated in Figure 5B. No significant difference was observed between the tumor AD and the contralateral values ( Figure 5A). The fractional anisotropy (FA) was significantly lower in the tumor on day 14 ( Figure 5C).
The MK was significantly higher in the tumor compared with the contralateral cortex (Z = 2.097, P = .036 for all time points) ( Figure 4B).
Median RK was also significantly higher in the tumor compared with the contralateral cortex on days 8 and 11 ( Figure 5E), whereas the tumor AK was not significantly different ( Figure 5D). Kurtosis fractional anisotropy (KFA) was significantly lower in the tumor compared with the contralateral cortex at all time points ( Figure 5F).

| Tumor vs. peritumoral edema
MD was significantly higher in the edema compared with the tumor and the contralateral cortex from day 8 (Z = 2.097, P = .036) ( Figure 4A). RD was also significantly higher in the edema compared with the tumor (Figure 5B), and AD was significantly higher only on day 14 ( Figure 5A). FA was significantly greater in the edema compared with the tumor ( Figure 5C). MK was significantly lower in the edema compared with the tumor (Z = 2.097, P = .036) ( Figure 4B). Significant differences were observed for AK at all time points (Figure 5D), and RK at days 8 and 11 ( Figure 5E).
KFA did not show any significant difference between the tumor and the peritumoral edema ( Figure 5F).
F I G U R E 6 A, FA and B, KFA maps of an ex vivo rat brain and C, corresponding 10X H&E staining. 20X magnification on D, the edematous region, E, the tumor edge, F, the tumor center and G, the contralateral cortex. The red arrows indicate the tumor on the FA and KFA maps

| Longitudinal changes in imaging parameters
The Friedman test showed no significant changes in tumor MD (χ 2 = 3, df = 2, P = .22) or MK (χ 2 = 1.33, df = 2, P = .51) values with time as tumor growth occurred. None of the other diffusivity and kurtosis parameters displayed any significant change with time and tumor growth.

| Ex vivo diffusivity and kurtosis
Ex vivo MRI scans and corresponding H&E slices of a representative brain are shown in Figure 6. A reduced FA was observed in the necrotic center of the tumor and in the tumor surroundings ( Figure 6A, Table 1). KFA followed the same trend ( Figure 6B, Table 1). Comparing the histological section with the similar slice section on MRI demonstrated a dense tumor (the visual appearance of higher staining reflecting increased cell density) on H&E staining ( Figure 6F). The necrotic center was hollow due to the fixation and dehydration processes. The tumor edge seemed to have elevated cellular density compared with the contralateral cortex ( Figure 6E).

| DISCUSSION
In this study, we investigated the utility of diffusion kurtosis to probe intra-tumoral heterogeneity in a rat model of intracranial glioblastoma.
Although diffusion kurtosis demonstrated significant differences between the tumor, the peritumoral edema and the contralateral cortex from the early stage onwards, none of the parameters changed significantly as the tumor grew.
We observed an increased MD and MK in the tumor in comparison with the contralateral cortex. The increased MD may be due to increased extracellular diffusivity, or a significant increase in intracellular water diffusion due to cellular swelling. An increased mean diffusivity in tumors relative to the contralateral cortex was reported in rats with F98 and C6 glioma, [8][9][10]29 although some discrepancy exists, as another study reported decreased MD in C6 gliomas. 30 MRI diffusion parameters, such as MD and FA, have shown potential for predicting tumor grade, 31,32 treatment monitoring and prognosis. 33 The F98 tumors exhibited higher MK values compared with the contralateral cortex, similar to some human studies reporting higher MK in high grade tumors compared with lower grade glioma. 23,34 However, higher MK has also been associated with inflammation and glial activity in a rat model of traumatic brain injury. 19,35 Hempel et al 36 reported that MK was a robust parameter for WHO classification of human gliomas. In fact, the highest MK values were measured in IDH WT glioblastoma described by an increased cellularity, cellular heterogeneity, hemorrhage, necrosis and microvasculature proliferation, and the lowest MK values were observed in IDH mut because of their low cell density and homogeneity. 37 The high MK observed in the F98 glioma in our study might originate from the high cellular density of the tumoral rim and the heterogeneity of the necrotic core, which was verified by H&E staining, whereby very high cell density was observed in the tumor, and a slightly increased cell density was observed in the peritumoral area compared with the contralateral cortex.
RD and AD exhibited the same trend as the MD values in the tumor and the contralateral cortex. Previous studies also reported higher RD in F98, 9L and GBM22 rat tumors. 8,10 We observed that RK was higher in the tumor compared with the contralateral cortex on days 8 and 11 whereas AK was not significantly different between the tumor and normal brain.
A lower FA value was observed in the tumor compared with normal brain in our study suggesting a more chaotic cellular organization in the tumor. In an earlier study, higher FA in the tumor rim than in the tumor core was reported in the F98 model. 9 As the tumor size increased, the necrotic core grew to become the major part of the tumor VOI by day 14, thereby contributing predominantly to the whole tumor diffusion anisotropy measurements in our study. Similar to our observations, increases in FA have been reported in human tumors from grade II to IV gliomas. 31 Lower KFA from the F98 tumor, especially from the necrotic center, indicates a much lower degree of tissue organization. KFA, which represents the anisotropy of the kurtosis tensor, has been recently proposed as a useful microstructural contrast. 38,39 Although this metric is more appropriate for WM analysis in the case of several crossing fiber orientations in the same voxel, it also seems to be of interest for gray matter (GM) microstructure, as elevated KFA was observed in tissues with low anisotropy such as the thalamus and lenticular nucleus, where the cells are organized in oriented structures (eg, lamina, nuclei). 39 The variability in normal brain VOI parameters ( Figure 5) was larger than expected, especially in the FA, RK and KFA values. The fact that this variability is not observed in all the parameters suggests that there might be some variability in the selection of the VOI, leading to different GM/WM ratios, and that FA, KFA and RK are probably more sensitive to these subtle alterations than the other DTI and DKI parameters. We observed decreased KFA in the tumoral tissue, suggesting a lower degree of overall tissue organization, which was noted in H&E stains showing high cellular density in the tumor with heterogeneity due to the necrotic cores.
The peritumoral edema displayed higher MD due to increased extracellular water. The increased water diffusion in all directions causes a significant decrease in diffusional kurtosis compared with the contralateral cortex, but also relative to the tumor. Furthermore, the peritumoral edema FA was always significantly higher than that of the tumor. An increased peritumoral edema FA and increased MD were also described in several F98 and 9 L rat glioma studies. [8][9][10] However, an increased FA and decreased ADC (MD) in the area surrounding the tumor was reported by Kim et al 10 and by Lope-Piedrafita et al in 9 L, F98 and C6 rat glioma, assumed to be caused by the compression of the surrounding cells to an oblate spheroid shape. 40 The increased diffusion anisotropy measured in the peritumoral region can be explained by the compression of the GM by the tumor mass, but also the infiltration of the tumor in the surrounding tissue. The H&E staining shows a higher cellular density in the edematous region. Furthermore, the cells seem to be more elongated in the region directly surrounding the tumor than in the contralateral cortex due to a tumoral mass effect, as suggested by Lope-Piedrafita et al. 40 It seems that our F98 glioma model is not only highly infiltrative but also demonstrates a mass effect on the nearby tissue.
DKI provides dimensionless metrics on the deviation of the probabilistic water displacement from a Gaussian distribution and has been proposed to better characterize tumor heterogeneity than standard DTI parameters in several pathological conditions. 24 In contrast to our hypothesis, we did not observe any temporal evolution in DKI parameters with tumor growth, as the tumor tissue clearly became more heterogenous with time, with increased necrotic areas. An increase in ADC over time has been reported by Letourneur et al 29 in a rat model with C6 glioma.
However, no changes in ADC were observed in F98 tumors 29 or 9 L tumors. 41 Our study did not demonstrate any better sensitivity in identifying tumor tissue from healthy brain with DKI compared with DTI at all imaging time points. Our initial hypothesis was that as the tumor grows, the increased microstructural heterogeneity due to hypoxia and necrosis would be quantifiable using DKI parameters. The lack of significant changes may either be due to tumor biology or due to limitations of the DKI technique. Initially, the DKI data were acquired at 9.4 T using two nonzero b-values that should theoretically allow kurtosis calculation in our model, 38,42 but only two b-values may not have provided sufficient sensitivity in measuring early microstructural changes, as suggested by other reports which used several low and high b-value combinations. 15,17 The use of the b = 0 values for both the DTI and DKI analysis led to some contributions from the fast (vascular) components of water diffusion due to the intravoxel incoherent motion (IVIM) effect. However, since the IVIM effect would have impacted both the DTI and DKI measurements, we believe that the IVIM effects would have cancelled out while comparing the two (DTI vs. DKI) for assessing tumor tissue heterogeneity. Additionally, although most human DKI studies are performed using 30 diffusionencoding directions, based on the recommendations of the DKE software, we used 15 diffusion directions in our study as the best compromise between SNR and acquisition time. However, we do not believe that the reduced number of diffusion direction impacts on the DKI fitting as it has been reported that DKI parameters can be calculated using a minimum of 15 diffusion directions. 15,16 Preclinical DKI studies have been reported with 15 directions in a rat model of stroke at 4.7 T 43 and 20 directions in diabetic rats at 7 T. 21 In fact, Lätt et al demonstrated that even six directions are sufficient to reach a good estimate of diffusion kurtosis in human MS at 3 T. 44 Another probable reason for not observing any change in kurtosis with tumor growth could be that the tumor was already highly heterogeneous (microstructurally) at the earliest imaging time point. In fact, a necrotic core was observed on the anatomical scans and parametric maps in all of the tumors from day 11 postimplantation (the second imaging time point). It is possible that the subsequent changes in tumor heterogeneity were not substantial enough to be detected with diffusion kurtosis MRI. The use of complementary imaging techniques can further aid in assessing the cellular swelling and extracellular matrix alterations such as time-dependent DTI, which was used in tumor models to separate intra-and extracellular water diffusion. 45,46 Alternatively, a slower growing tumor model could be used, or a treatment paradigm that substantially alters the tissue microstructure by induction of cell death.
In conclusion, an increased diffusional kurtosis in F98 tumors, and a decrease in the peritumoral edema, were observed compared with normal brain, although no changes in DKI parameters were noted as the tumor grew, indicating that this technique may not be able to observe the microstructural tumor heterogeneity in the F98 model.