Automated quantification of brain PET in PET/CT using deep learning-based CT-to-MR translation: a feasibility study.

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Quantitative analysis of brain PET/CT scans typically requires MRI-derived regions of interest (ROIs).
• Challenges arise from the frequent unavailability of paired PET/CT and MRI scans.
• Misalignment issues occur when PET/CT and MRI acquisition times differ significantly.

Purpose of the Study:

• To develop a deep learning framework for translating CT images from PET/CT scans into synthetic MR images (MR_SYN).
• To enable automated quantitative regional analysis of brain PET/CT images using MR_SYN-derived segmentation.
• To overcome limitations associated with the absence or temporal mismatch of MRI scans.

Main Methods:

• A retrospective study included 139 subjects with [¹⁸F]FBB PET/CT and T1-weighted MRI.
• A U-Net-like model was trained for CT-to-MR_SYN image translation.
• A separate model performed segmentation of MR_SYN into 95 regions for quantitative standardized uptake value ratio (SUVr) calculation.
• Evaluation involved Structural Similarity Index Measure (SSIM) for MR_SYN quality and Dice Similarity Coefficient (DSC) for segmentation accuracy.

Main Results:

• The synthetic MR images (MR_SYN) achieved a high mean SSIM of 0.974 ± 0.005 compared to ground-truth MR (MR_GT).
• MR_SYN-based segmentation yielded a mean DSC of 0.733 across 95 regions.
• Quantitative SUVr measurements derived from MR_SYN showed no statistical significance compared to MR_GT, except for the precuneus.

Conclusions:

• A deep learning framework was successfully demonstrated for automated regional brain analysis in PET/CT using synthetic MR images.
• This method offers a valuable alternative for quantitative PET/CT analysis in patients unable to undergo MRI scans.
• The framework facilitates robust quantitative imaging in scenarios with limited or no MRI availability.