MRI-based attenuation correction for brain PET/MRI based on anatomic signature and machine learning

Area of Science:

• Medical Imaging
• Machine Learning
• Radiophysics

Background:

• Accurate attenuation maps are crucial for PET/MRI but challenging due to MRI signal limitations.
• Existing methods struggle with bone/air interfaces and MRI's lack of direct attenuation information.

Purpose of the Study:

• To develop and validate a learning-based method for generating patient-specific CT attenuation maps from routine T1-weighted MRI.
• To enable accurate attenuation correction for brain PET imaging in PET/MRI scanners.

Main Methods:

• A machine learning approach using alternating random forests within an iterative refinement model.
• Inclusion of anatomical feature selection during training and prediction.
• Retrospective analysis of 17 patients with PET/CT and MR brain scans.

Main Results:

• Generated pseudo-CT (PCT) images demonstrated high accuracy with mean average error of 66.1 ± 8.5 HU and DSC > 0.85 for key tissues.
• PET images corrected using PCT showed excellent agreement with CT-corrected PET images (mean difference < 4% in VOIs).
• Quantitative analysis revealed high correlation (0.989 ± 0.017) and low differences (2.5% in active areas) between PCT and CT corrected PET.

Conclusions:

• The developed learning-based method effectively generates CT images from MRI for PET attenuation correction.
• PET imaging using PCT-derived attenuation maps achieves high quantitative equivalence to PET/CT.
• This approach offers a viable solution for accurate PET attenuation correction in MRI-only or PET/MRI settings.