Repeatability of derived parameters from histograms following non-Gaussian diffusion modelling of diffusion-weighted imaging in a paediatric oncological cohort

Abstract

Objectives: To examine repeatability of parameters derived from non-Gaussian diffusion models in data acquired in children with solid tumours. Methods: Paediatric patients (<16 years, n = 17) were scanned twice, 24 h apart, using DWI (6 b-values, 0–1000 mm−2 s) at 1.5 T in a prospective study. Tumour ROIs were drawn (3 slices) and all data fitted using IVIM, stretched exponential, and kurtosis models; percentage coefficients of variation (CV) calculated for each parameter at all ROI histogram centiles, including the medians. Results: The values for ADC, D, DDCα, α, and DDCK gave CV < 10 % down to the 5th centile, with sharp CV increases below 5th and above 95th centile. K, f, and D* showed increased CV (>30 %) over the histogram. ADC, D, DDCα, and DDCK were strongly correlated (ρ > 0.9), DDCα and α were not correlated (ρ = 0.083). Conclusion: Perfusion- and kurtosis-related parameters displayed larger, more variable CV across the histogram, indicating observed clinical changes outside of D/DDC in these models should be interpreted with caution. Centiles below 5th for all parameters show high CV and are unreliable as diffusion metrics. The stretched exponential model behaved well for both DDCα and α, making it a strong candidate for modelling multiple-b-value diffusion imaging data. Key Points: • ADC has good repeatability as low 5th centile of the histogram distribution. • High CV was observed for all parameters at extremes of histogram. • Parameters from the stretched exponential model showed low coefficients of variation. • The median ADC, D, DDCα, and DDCKare highly correlated and repeatable. • Perfusion/kurtosis parameters showed high CV variations across their histogram distributions.