Neural Network Methods Enhance Test-Retest Repeatability of IVIM Parameter Estimation in Pancreatic Imaging

Motivation: IVIM is valuable for differentiating pancreatic diseases, but parameter reliability depends significantly on algorithm choice. Robust methods are essential to enhance repeatability, ensuring accurate assessments that support clinical decisions. Goal(s): To assess whether neural-network-based approaches enhance the repeatability of IVIM parameter fitting in the pancreas compared to traditional nonlinear least-squares methods. Approach: Test-retest diffusion-weighted MRI was acquired from eight healthy subjects and analyzed for repeatability (CV) and goodness of fit (R2) using two nonlinear least-squares methods and three neural networks. Results: Neural-network approaches, especially IVIM-Morph and Super-IVIM-DC, achieved significantly lower CVs than conventional methods with comparable R2, improving repeatability without compromising fit quality. Impact: This study demonstrates that neural network-based methods, specifically the motion-robust IVIM-Morph and SUPER-IVIM-DC for limited b-values, improve test-retest repeatability in IVIM parameter estimation for pancreatic imaging, supporting better assessment of pancreatic conditions and enhancing diagnosis, treatment planning, and patient outcomes.