Uncertainty estimates in pharmacokinetic modelling of DCE-MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Quantitative MRI

Background:

• Dynamic contrast-enhanced MRI (DCE-MRI) quantifies tissue perfusion but faces challenges in accuracy due to noisy data and complex modeling.
• Conventional methods like non-linear least squares (NLLS) fitting produce noisy parameter maps.
• Deep learning models offer smoother maps but lack reliable uncertainty quantification, potentially misleading clinicians.

Purpose of the Study:

• To implement and evaluate an ensemble of mean-variance estimation (MVE) neural networks for DCE-MRI.
• To quantify perfusion parameters and associated aleatoric and epistemic uncertainties.
• To compare MVE's performance against NLLS and physics-informed neural networks (PINNs) for uncertainty estimation.

Main Methods:

• Developed an ensemble of MVE neural networks for DCE-MRI perfusion quantification.
• Implemented conventional covariance matrix-based uncertainty estimation for NLLS and PINNs.
• Compared MVE with NLLS and PINNs using simulations and in vivo data, focusing on perfusion accuracy and uncertainty estimation.

Main Results:

• MVE demonstrated superior accuracy in both perfusion and uncertainty estimates compared to NLLS and PINNs in simulations.
• MVE's aleatoric uncertainty closely correlated with actual errors, unlike NLLS and PINNs which tended to overestimate.
• In vivo, MVE generated smoother, more reliable uncertainty maps, particularly in liver regions, outperforming NLLS and PINNs.

Conclusions:

• MVE enhances quantitative DCE-MRI by providing robust perfusion parameters and uncertainty estimates.
• This approach increases the reliability of AI-driven MRI analysis, fostering greater clinical confidence.
• MVE facilitates the clinical translation of advanced quantitative MRI techniques.