Voxel-wise uncertainty in CT substitute derived from MRI

Area of Science:

• Medical Imaging
• Radiology
• Image Reconstruction

Background:

• Substitute CT (s-CT) images can be generated from MRI using ultrashort echo time (UTE) and T2-weighted sequences with Gaussian mixture regression (GMR).
• Quantifying the uncertainty in these s-CT reconstructions is crucial for clinical application.

Purpose of the Study:

• To analyze uncertainties in the GMR model for s-CT image generation.
• To determine the information contributions of individual MRI sequences in the GMR model.

Main Methods:

• Derived an analytical expression for voxel-wise expected absolute deviation (EAD) in s-CT images.
• Evaluated the EAD against the true mean absolute prediction deviation (MAPD) in 14 patients.
• Investigated the impact of excluding individual MR sequences on s-CT quality.

Main Results:

• The highest EAD was observed at air-soft tissue and bone-soft tissue interfaces.
• The derived EAD accurately predicted MAPD in high-contrast and low-contrast regions.
• Two MR sequences were found to be redundant for s-CT generation, including a non-UTE sequence.

Conclusions:

• The developed uncertainty estimation method reliably predicts voxel-wise MAPD in s-CT images.
• A previously used non-UTE sequence is redundant for GMR-based s-CT generation, simplifying the model.