Deep Learning-Based Parameter Mapping with Uncertainty Estimation for Fat Quantification using Accelerated...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Quantitative MRI

Background:

• Deep learning (DL) is used for artifact removal in undersampled MRI and signal fitting in quantitative MRI.
• These tasks are typically handled separately, limiting the exploitation of shared information.
• Accelerated MRI protocols are crucial for reducing scan times, especially in applications like liver fat quantification.

Purpose of the Study:

• To develop a unified, two-stage deep learning framework for artifact removal and signal fitting in quantitative MRI.
• To estimate pixel-wise uncertainty levels within the quantitative MRI data.
• To improve the efficiency and accuracy of liver fat quantification using accelerated free-breathing radial MRI.

Main Methods:

• A novel two-stage deep learning framework was proposed, integrating artifact removal and signal fitting.
• The framework was applied to accelerated free-breathing radial MRI data for liver fat quantification.
• Pixel-wise uncertainty estimation was incorporated into the framework.

Main Results:

• The proposed framework achieved high image quality from undersampled radial data.
• Accurate liver fat quantification was demonstrated with reduced computational time (<100 ms/slice).
• The framework successfully detected uncertainty arising from noisy input data.

Conclusions:

• The integrated framework effectively addresses artifact removal and signal fitting in quantitative MRI.
• The method enables significant acceleration (3-fold) with a scan time under 1 minute.
• This approach enhances the feasibility of rapid and reliable liver fat quantification in clinical settings.