The Intravoxel Incoherent Motion (IVIM) model has recently gained more interest with its derivation of diffusion and perfusion related parameters (pseudodiffusion, D* and perfusion fraction, f).¹ The IVIM method has largely been accepted in the abdomen, ² but applications in the brain and in paediatrics have not been explored to a great extent. ³ Therefore, the aim of this study was to evaluate the appropriateness of the fitting methods (Fig. 1) in terms of the degree of bi-exponential behaviour and the signal-to-noise ratio (SNR) with IVIM parameters derived from normal brain and tumours. A comparison to DSC-MRI derived perfusion parameters was used to verify the physical origins of the IVIM parameters.

The IVIM and DSC data analysis was computed using in-house Python software (Anaconda, Continuum Analytics, v.2.7.10) and the statistical analysis in SPSS Statistics (IBM, v.22). For IVIM model simulations and patient data, non-linear least-squares minimisation was performed using the Levenberg-Marquardt algorithm. The bi-exponential fitting methods varied from a non-constrained simultaneous fitting to a more constrained step-wise fitting (Fig.1). The IVIM grey matter (GM) and tumour model data were created using IVIM parameters derived from the relevant ROIs and with the same b-value distribution as used for the patient imaging. Random Gaussian noise was added to the model data to mimic SNR levels of 10-70.

Patient data (6 brain tumour patients aged 1-10) was acquired on a Philips Achieva 3T TX MRI scanner at the Birmingham Children’s Hospital using 32-multichannel receiver head coil. The DW-MRI sequence used a sensitivity-encoded (SENSE) approach with the following parameters: 11 exponentially spaced b-values (0-1000 s mm^-2) in three orthogonal directions, TR/TE 4,000/91 ms, FOV 240x240, acquisition matrix 96x96, 30 slices with thickness of 3.5 mm and in-plane resolution 2.5x2.5 mm². The scan duration was 2.12 minutes. The DSC-MRI used a protocol described previously, as well as the computation of cerebral blood flow (CBF), cerebral blood volume (CBV) and mean transit time (MTT). ⁴ ROIs were drawn manually on co-registered T2-weighted images for healthy-appearing GM and tumours and transferred to the perfusion maps for correlation analysis (Spearman correlation coefficient, ρ).

The data simulation results are shown in Fig.2 and Table 1. The accuracy was measured using the theoretical IVIM value and the mean value from 1000 data iterations and the precision in terms of coefficient of variation. In the realistic SNR_b1000 of 30 ³, the constrained 1-parameter fitting was found to produce the overall most accurate and precise IVIM parameters f and D* with average error/CV% 11.4/77.6% and 37.6/69.9 % in the low-perfusion regime and 16.4/35.8% and 3.33%/39.7% in the high-perfusion regime, respectively.

Significant voxel-wise correlations were observed in the GM between the IVIM-f and DSC-CBV (ρ = 0.271-0.588, P <0.01), IVIM-fD* and DSC-CBF (ρ = 0.302-0.543, P <0.01), respectively. The correlation between the IVIM-D* and DSC-MTT was not significant (ρ = 0.024-0.162, P >0.05). The voxel-wise correlations in the tumours were found more variable with only a few cases showing any significant correlation (f-CBV ρ = 0.024-0.154, D*-MTT ρ = 0.009-0.388, fD*-CBF ρ = 0.031-0.434). Relative mean (mean tumour/mean GM) correlations were also computed for the perfusion parameters (Fig.4). Significant correlation was observed between the IVIM-f and DSC-CBV parameters, using both 1-and 2-parameter fitting. Strong correlations were also observed between the IVIM-fD* and DSC-CBF.

The simulations suggest that the IVIM-f is robust enough to provide reliable results for low-perfused tumours and GM. The IVIM-D* was found to have limited robustness for the tumour model, with noticeable higher variance compared to IVIM-f (Fig.2). Overall, the 1-parameter fitting was found to be the most robust in terms of accuracy and precision, with the smallest number of outliers (Table 1).

The strong correlation of IVIM-f and DSC-CBV confirms the perfusion related origins of the f-parameter and further confirms the robustness of the 1-parameter fitting (Fig.4 and 5). The correlations between IVIM-D*, fD* and DSC-CBF, MTT, respectively, were likely reduced due to the limited robustness of IVIM-D* fitting as observed in the simulations. The IVIM-D* and DSC-MTT were found to correlate for the younger cohort (ρ = 0.800, P <0.05), but an outlier case (age 10) was found to reduce the correlation significantly. A larger cohort with a range of ages is required to further investigate the effect of age on perfusion.

The IVIM-f parameter using the 1-parameter fitting was found to provide a robust measure of CBV in GM and low-perfused tumours in this paediatric cohort. With anti-angiogenic treatment being increasingly used in these brain tumours, IVIM-f could provide a contrast-free biomarker for assessing these tumours.