This paper highlights the difficulty of accurately estimating intravoxel incoherent motion (IVIM) parameters in diffusion-weighted MRI (DMR) due to the instability of the inverse problem and its high sensitivity to noise, especially in the perfusion compartment. To address this, we propose a probabilistic deep learning framework based on deep ensembles (DE) of Mixture Density Networks (MDNs). This framework estimates the overall prediction uncertainty and decomposes it into alleatoric uncertainty (AU) and epistemic uncertainty (EU). The proposed method is evaluated against non-probabilistic neural networks, Bayesian fitting methods, and probabilistic networks with single Gaussian parameterization. Supervised learning is performed using synthetic data, and evaluations are performed on simulated and in vivo datasets. The reliability of the quantified uncertainty is assessed using calibration curves, sharpness of the output distribution, and continuous rank probability scores (CRPS). MDNs produced more corrected and sharper predictive distributions for the diffusion coefficient D and fraction f parameters, but slight overconfidence was observed for the pseudodiffusion coefficient D . The robust coefficient variability (RCV) indicated smoother in vivo estimates of D using MDNs compared to the Gaussian model. Despite training data covering the expected physiological range, the high in vivo EU suggests a discrepancy with actual acquisition conditions and highlights the importance of EU integration allowed by DE. Overall, this paper presents a comprehensive framework for IVIM fitting via uncertainty quantification, which allows for the identification and interpretation of unreliable estimates. The proposed approach can also be applied to other physical model fitting with appropriate architecture and simulation adjustments.
