Optimizing IVIM measurement on magnetic resonance linear accelerator using multi-b-value acquisition and physics-informed neural network

Background: Accurate quantification of intravoxel incoherent motion (IVIM) parameters on the magnetic resonance linear accelerator (MR-Linac) is challenging due to low signal-to-noise ratio and time constraints. Purpose: This study aimed to evaluate the impact of b -value acquisition strategies and estimation algorithms on IVIM parameter measurement to guide reliable IVIM quantification on the MR-Linac. Methods: A digital phantom with high number of signals averaged (high-NSA) and multi- b -value signals was constructed to evaluate estimation accuracy using root mean square error (RMSE). Ten healthy volunteers underwent four repeated diffusion-weighted imaging (DWI) scans, consisting of two high-NSA DWIs and two multi- b -value DWIs to assess repeatability via the within-subject coefficient of variation (wCV). IVIM parameters from high-NSA DWI were estimated using Segment fitting, while multi- b -value DWI data were processed using least squares (LS), Bayesian estimation, and physics-informed neural network (PINN). Results: All IVIM parameters from multi- b -value DWI exhibited lower wCVs than those from high-NSA DWI, except for the LS-derived diffusion coefficient ( D ). Multi- b -value DWI combined with PINN achieved rapid and highly repeatable estimation. In the digital phantom, high-NSA DWI with Segment fitting yielded the highest RMSE for the pseudo-diffusion coefficient ( D ∗ ). LS and Bayesian methods resulted in higher D RMSEs and required longer fitting times. In contrast, multi- b -value DWI + PINN achieved the lowest RMSEs across all parameters with minimal computational cost. Conclusion: The combination of multi- b -value acquisition and PINN provided a robust and accurate framework for IVIM parameter estimation on the MR-Linac and showed potential for integration into MR-Linac quantitative imaging workflows.