Neural networks improve repeatability of intravoxel incoherent motion (IVIM) parameter estimation in pancreatic diffusion-weighted MRI.

Intravoxel incoherent motion (IVIM) analysis in diffusion-weighted MRI (DWI-MRI) shows potential for characterizing pancreatic tissue, but its clinical application remains limited by sensitivity to fitting algorithms. This study assessed the repeatability of neural network (NN)-based IVIM fitting versus classical nonlinear least-squares in pancreatic DWI. The repeatability cohort included ten healthy volunteers and two type 1 diabetes (T1D) individuals, each scanned twice; the glucose-response cohort included three T1D participants and three healthy controls scanned pre- and post-oral glucose. Diffusion data were acquired at 13 b-values (0-1200 s/mm[Formula: see text]). Four NN-based methods (IVIM-NET, SUPER-IVIM-DC, U-Net, IVIM-MORPH) were compared with two classical approaches (SLS, SLS-TRF) using full (13-point) and reduced (7-point) protocols. Repeatability was quantified using within-subject coefficient of variation (wCV) and Bland-Altman analysis. All NN-based methods significantly improved test-retest repeatability of the perfusion fraction ƒ compared with classical approaches ([Formula: see text]), with SUPER-IVIM-DC showing the lowest wCV and IVIM-MORPH offering balanced performance across parameters. A reduced protocol shortened scan time through fewer acquisitions while maintaining or improving repeatability compared to the full protocol. Preliminary glucose-response results show both NN and classical IVIM analyses detect physiologically relevant changes. Parameter estimates varied across NN architectures, requiring further validation to establish accuracy.