Three sets of parameters were used to simulate noisy IVIM signals in different perfusion regimes (from literature {D,f,D*}-Low:{1e-3mm²/s, 0.05, 1e-2mm²/s}; Medium:{15e-4mm²/s, 0.30, 15e-3mm²/s}; High:{1e-3mm²/s, 0.30, 6e-2mm²/s}). A segmented fit was performed for b-low ([0 190]s/mm², step:10s/mm²) and b-high ([300 1400]s/mm², step:100s/mm²) b-values ranges. The optimal b-values set was found through an optimization procedure: from thousands of sets of 13 random b-values, estimates errors were computed (σₜₒₜ as the sum of relative errors of IVIM parameters: σD, σD*, σf) and candidate sets were chosen as those for which σD ≤ 0.25 for b-high and σₜₒₜ ≤ 0.80 for b-low. The 13 values of the optimal sets (b-opt₁₃, one for each regime) were found as the most frequent b-values in b-high and b-low. Optimized sets of 7 b-values (b-opt₇) were obtained via backward elimination: the least frequent b-values of b-opt₁₃ were iteratively eliminated in b-low and b-high ranges to minimize σₜₒₜ.
Simulations were also performed using 7 b-values used in the clinical practice (b-im=[0 50 100 150 200 400 1000]s/mm²). Relative errors from different perfusion regimes and b-values sets (b-opt₁₃, b-opt₇ and b-im) were compared to evaluate differences between b-values configurations (Kruskall-Wallis test, α=0.05).
To evaluate the impact of optimization on patient data, IVIM maps were computed voxel wise through a segmented fit of diffusion MRI data acquired with b-im from patients affected by meningioma (n=9) and skull-base chordoma (n=11). The means of D, f and D* were computed over all patients in white matter (WM) and gross tumour volume (GTV) regions, and fed into simulations to compare b-opt₇ to b-im (Wilcoxon tests, α=0.05) and characterize tissue perfusion regimes.