Pixel-Wise Uncertainty Quantification of Accelerated MRI Reconstruction

Area of Science:

• Medical Imaging
• Machine Learning
• Image Reconstruction

Background:

• Accelerated magnetic resonance imaging (MRI) reduces scan times but can compromise image quality.
• Assessing the reliability of under-sampled MRI reconstructions is crucial for clinical applications.
• Existing methods often require ground-truth data, limiting their utility in real-world scenarios.

Purpose of the Study:

• To develop an automated method for evaluating the quality of under-sampled MRI reconstructions.
• To enable pixel-wise uncertainty quantification in accelerated MRI without relying on fully-sampled reference images.
• To identify unreliable regions in MRI reconstructions automatically.

Main Methods:

• A framework integrating conformal quantile regression with learning-based reconstruction was proposed.
• Statistically rigorous pixel-wise uncertainty intervals were estimated.
• An end-to-end Variational Network was employed for image reconstruction on the fastMRI dataset (acceleration factors 2-10).

Main Results:

• High agreement (Pearson correlation >90% at 4x acceleration) was observed between predicted uncertainty maps and actual reconstruction errors.
• The proposed method outperformed a simpler heuristic approach (residual magnitude).
• Qualitative analysis showed uncertainty maps accurately reflected reconstruction errors, aligning with artifacts and pathologies.

Conclusions:

• The developed framework allows for reconstruction quality assessment without ground-truth data.
• This work advances adaptive MRI protocols, balancing scan time and diagnostic reliability.
• Automated uncertainty quantification enhances the trustworthiness of accelerated MRI.