WE-C-217BCD-12: Irreversible Two-Tissue Compartment Model Fitting for Dynamic 18F- FDG PET: A Practical Comparison of...

Area of Science:

• Nuclear Medicine
• Radiochemistry
• Biophysics

Background:

• Dynamic Positron Emission Tomography (PET) analysis is vital for quantitative imaging.
• Fitting time-activity curves (TACs) to kinetic models is essential for extracting physiological parameters.
• Computational efficiency and parameter accuracy are key considerations in voxel-wise PET analysis.

Purpose of the Study:

• To compare the precision and accuracy of various algorithms for fitting dynamic 18F-FDG PET TACs to the irreversible two-tissue compartment model.
• To evaluate algorithm performance under measurement error and assess computational time (CPU).

Main Methods:

• Simulated dynamic 18F-FDG PET TACs were generated using a two-tissue compartment model with fixed biological parameters.
• Gaussian noise was incorporated to simulate measurement error.
• Algorithms evaluated included Genetic, Conjugate Gradient (CG), Gradient Descent (GD), Simulated Annealing (SA), Levenberg-Marquardt (LMQ), Gauss-Newton (GN), Limited-BFGS (L-BFGS), Patlak analysis, and Blomqvist linearization (BL).
• Parameter accuracy and precision were quantified using relative errors (REs).

Main Results:

• Gradient Descent and Simulated Annealing showed maximal REs >60% at typical noise levels.
• Blomqvist linearization and Patlak analysis yielded REs <20% and <4% respectively.
• Limited-BFGS, Gauss-Newton, and Levenberg-Marquardt achieved the highest fidelity, with rate-constant REs <2.5%.
• Non-iterative methods (Patlak, BL) were significantly faster (<0.05 ms) but less accurate.
• Combining BL with L-BFGS, GN, or LMQ (BL+L-BFGS, BL+GN, BL+LMQ) provided rapid convergence (<0.1 ms) with superior parameter bias.

Conclusions:

• The selection of a fitting algorithm significantly impacts the accuracy of kinetic parameters and computation time in dynamic PET.
• Recommended methods, BL+L-BFGS, BL+GN, or BL+LMQ, offer a superior balance of accuracy and speed compared to commonly used methods like Patlak or LMQ alone.
• These hybrid approaches provide a practical solution for efficient and accurate voxel-wise dynamic PET analysis.