Uncertainty Quantification in Deep MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) models offer potential for accelerated MRI reconstruction but introduce uncertainty.
• Undersampled MRI scans and DL model opacity challenge diagnostic accuracy.

Purpose of the Study:

• To quantify uncertainty in DL-based MRI image reconstruction.
• To develop a probabilistic framework for assessing reconstruction accuracy.

Main Methods:

• Utilized variational autoencoders (VAEs) for probabilistic reconstruction, encoding acquisition uncertainty.
• Employed Monte-Carlo sampling for pixel variance maps and Stein's Unbiased Risk Estimator (SURE) for bias estimation.
• Evaluated reconstruction performance on Knee MRI data using adversarial and pixel-wise losses with unrolled recurrent networks.

Main Results:

• Adversarial training losses increased reconstruction uncertainty.
• Recurrent unrolled network architectures decreased prediction uncertainty and risk.
• Probabilistic VAEs enabled generation of pixel variance maps for uncertainty quantification.

Conclusions:

• DL reconstruction uncertainty can be quantified using probabilistic VAEs and SURE.
• Recurrent architectures show promise in reducing uncertainty for more reliable MRI.
• Understanding and mitigating DL uncertainty is crucial for clinical diagnostic and therapeutic applications.