Predictive uncertainty in deep learning-based MR image reconstruction using deep ensembles: Evaluation on the fastMRI...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) models for MR image reconstruction risk generating unrealistic artifacts or missing pathologies.
• Current DL methods are often task-agnostic and not robust to domain shifts, potentially leading to undetected errors.
• Quantifying uncertainty in DL reconstructions is crucial for assessing robustness and reliability.

Purpose of the Study:

• To develop and evaluate a method for pixel-wise predictive uncertainty estimation in DL-based MR image reconstruction.
• To investigate the impact of domain shifts and varying network architectures on reconstruction uncertainty.
• To differentiate between aleatoric (data) and epistemic (model) uncertainty.

Main Methods:

• Proposed a strategy combining deep ensembles for epistemic uncertainty and nonnegative log-likelihood loss for aleatoric uncertainty.
• Integrated this uncertainty estimation with conventional loss terms for DL reconstruction.
• Evaluated five different DL architectures on the fastMRI database, testing with in-distributional and out-of-distributional data (varying undersampling, contrast, orientation, anatomy, pathology).

Main Results:

• The proposed uncertainty measure effectively captured pixel-wise predictive uncertainty, correlating well with normalized mean squared error.
• Uncertainty was localized to aliased anatomies and regions with abnormal signal intensity (hyper/hypointense).
• The method successfully detected shifts in disease prevalence and revealed distinct uncertainty patterns across different network architectures.

Conclusions:

• The developed approach enables robust estimation of both aleatoric and epistemic uncertainty in DL-based MR reconstruction.
• Provides an interpretable, pixel-level examination of predictive uncertainty, enhancing model reliability.
• Facilitates a deeper understanding of DL model behavior under various conditions and potential domain shifts.