Physics-Informed CNN Ensemble for Improved Tissue Perfusion and Diffusion Estimation with Uncertainty Quantification in IVIM MRI

This study explores a physics-informed convolutional neural network (CNN) ensemble for Intravoxel Incoherent Motion (IVIM) parameters estimation in diffusion-weighted MRI (DW-MRI). A U-Net was trained on simulated phantoms under different noise conditions and evaluated using Median Absolute Error and Robust Coefficient of Variation. The ensemble approach improved robustness and enabled uncertainty quantification through variance maps. When tested on in-vivo mouse brain images, the CNN estimates aligned well with Bayesian results, particularly for D and D* parameters. The findings suggest that ensembling improves IVIM accuracy and reliability while also enabling uncertainty quantification.