PixCUE: Joint Uncertainty Estimation and Image Reconstruction in MRI using Deep Pixel Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning (DL) models excel at accelerated MRI reconstruction by utilizing latent data representations.
• Key challenges in DL for MRI include inherent uncertainties from k-space undersampling and the opaque nature of DL models.
• Accurate uncertainty estimation is critical for reliable DL-based MRI reconstruction.

Purpose of the Study:

• To develop an efficient method for uncertainty estimation in DL-based MRI reconstruction.
• To introduce PixCUE (Pixel Classification Uncertainty Estimation) for simultaneous image reconstruction and uncertainty mapping.
• To validate PixCUE's performance against reconstruction errors and conventional uncertainty estimation techniques.

Main Methods:

• Proposed PixCUE, a novel pixel classification framework for uncertainty estimation in DL MRI reconstruction.
• PixCUE performs image reconstruction and uncertainty map generation in a single forward pass.
• Validated PixCUE's uncertainty maps against reconstruction errors across various MR sequences and adversarial conditions.

Main Results:

• PixCUE-generated uncertainty maps strongly correlate with reconstruction errors (NMSE, PSNR, SSIM).
• Established an empirical relationship between PixCUE uncertainty estimations and standard reconstruction metrics.
• Demonstrated a significant correlation between PixCUE uncertainty estimates and conventional Monte Carlo (MC) methods.
• PixCUE reliably estimates uncertainty with minimal additional computational cost.

Conclusions:

• PixCUE offers an efficient and reliable approach for uncertainty estimation in DL-based accelerated MRI reconstruction.
• The method provides valuable uncertainty maps in a single forward pass, reducing computational burden.
• PixCUE's strong correlation with established metrics and MC methods validates its efficacy for clinical applications.