Deep Learning-Based Precontrast CT Parcellation for MRI-Free Brain Amyloid PET Quantification.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate amyloid quantification in 18F-FBB PET/CT is crucial for diagnosing and monitoring neurodegenerative diseases.
• High-resolution MRI is typically required for precise brain region parcellation, which is not always available or feasible.
• Developing alternative methods for brain segmentation using more widely available imaging modalities like CT is essential.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for brain region parcellation using CT data from PET/CT scans.
• To enable accurate amyloid quantification in 18F-FBB PET/CT without reliance on high-resolution MRI.
• To assess the agreement between CT-based DL parcellation and traditional MRI-based methods for amyloid quantification.

Main Methods:

• A retrospective dataset of 226 individuals (PET/CT and MRI pairs) was utilized, split into training/validation (60%) and test (40%) sets.
• Three UNet models were independently trained for multiplanar brain parcellation on CT data and subsequently ensembled.
• Amyloid load was quantified across 46 volumes of interest (VOIs) using standardized uptake value ratios (SUVRs), with comparisons made using Dice similarity coefficients, linear regression, and intraclass correlation coefficients.

Main Results:

• The DL-based CT parcellation achieved a mean Dice similarity coefficient of 0.80 for all 46 VOIs, with higher scores for subcortical (0.83) and lower for cortical/limbic (0.72) regions.
• Comparisons of regional and global SUVRs between CT-based DL and MRI-based methods showed excellent agreement (R² ≧ 0.976, ICC ≧ 0.988).
• Both methods demonstrated a consistent increase in global SUVR with increasing Clinical Dementia Rating (CDR) scores, confirming the model's clinical relevance.

Conclusions:

• The developed deep learning model for brain parcellation using CT data demonstrates strong agreement with MRI-based methods.
• This CT-based approach offers a viable alternative for accurate amyloid quantification in 18F-FBB PET/CT, reducing the need for concurrent MRI.
• The findings support the potential of DL-driven CT parcellation to improve accessibility and efficiency in neurodegenerative disease imaging.