Validation of a Deep-Learning Coregistration Framework for Long-Axial-Field-of-View PET/CT Using...

Area of Science:

• Medical Imaging
• Radiochemistry
• Artificial Intelligence

Background:

• Long-axial-field-of-view (LAFOV) PET/CT systems enhance whole-body imaging sensitivity and coverage.
• PET/CT misalignment can lead to artifacts and quantification errors, especially with reduced scan times and doses.
• Ultra-low-dose CT (ULD-CT) is increasingly used, posing challenges for accurate image registration.

Purpose of the Study:

• To evaluate a deep-learning coregistration framework for PET/CT alignment in LAFOV systems.
• To assess the framework's robustness with low-dose PET and ULD-CT across different radiotracers.
• To validate the framework's performance using quantitative metrics and SUV accuracy.

Main Methods:

• Included 63 scans with four tracers (⁸⁹Zr-trastuzumab, ¹⁵O-H₂O, ¹⁸F-MC225, ¹⁸F-FDG) on a LAFOV PET/CT system.
• Compared coregistration of standard CT to PET (rCT) versus low-count PET (LC-rCT) and ULD-CT to PET (rULD-CT) versus low-dose CT (rLD-CT).
• Utilized Dice similarity, Jaccard similarity, Hausdorff distance, and average surface distance for evaluation; assessed SUV for PET accuracy.

Main Results:

• The deep-learning framework demonstrated robust performance across tracers, even with reduced PET counts and ULD-CT.
• Quantitative metrics showed consistent alignment accuracy between standard and reduced-dose/count scenarios.
• Standardized Uptake Values (SUVs) remained consistent, confirming the quantitative accuracy of the coregistration approach.

Conclusions:

• The deep-learning coregistration framework is effective for PET/CT alignment in LAFOV systems, even with low-dose PET and ULD-CT.
• This approach supports the feasibility of reduced radiation exposure protocols without compromising image registration quality.
• Findings validate the framework's generalizability and robustness for clinical and research applications using various tracers.