Uncertainty modelling in deep learning for safer neuroimage enhancement: Demonstration in diffusion MRI

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Deep learning (DL) excels in medical image enhancement but often uses deterministic models, ignoring crucial uncertainty.
• Existing DL methods for super-resolution and synthesis overlook inherent data and model uncertainties.

Purpose of the Study:

• To develop and demonstrate methods for characterizing and quantifying uncertainty in DL-based medical image enhancement.
• To improve the safety and reliability of DL models in diffusion MRI super-resolution.

Main Methods:

• Proposed methods to account for intrinsic uncertainty via heteroscedastic noise modeling and parameter uncertainty via approximate Bayesian inference.
• Integrated intrinsic and parameter uncertainty to quantify predictive uncertainty in output images.
• Developed a method to propagate predictive uncertainty to derived scalar parameters (e.g., mean diffusivity, fractional anisotropy).

Main Results:

• Uncertainty modeling improved predictive performance on out-of-distribution datasets.
• Predictive uncertainty correlated with reconstruction errors, enabling detection of predictive failures.
• Uncertainty decomposition provided explanations for model performance by quantifying sources of uncertainty.

Conclusions:

• Uncertainty quantification enhances DL safety in medical imaging, improving performance and reliability.
• The developed methods enable subject-specific, voxel-wise risk assessment for super-resolved images.
• The approach offers insights into model behavior and extends to other imaging modalities and applications.