Premium
WE‐C‐217BCD‐12: Irreversible Two‐Tissue Compartment Model Fitting for Dynamic 18F‐ FDG PET: A Practical Comparison of Methods Using Simulated Time‐ Activity Data
Author(s) -
McDermott S,
Yan D
Publication year - 2012
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4736128
Subject(s) - broyden–fletcher–goldfarb–shanno algorithm , conjugate gradient method , algorithm , mathematics , voxel , computer science , artificial intelligence , computer network , asynchronous communication
Purpose: To compare techniques of fitting simulated Dynamic 18F‐FDG PET time‐activity curves to the irreversible two‐tissue compartment model. The precision and accuracy of various algorithms were assessed in the presence of measurement error and with a practical focus on CPU time. Methods: Dynamic PET analysis is often applied to multiple individual voxels, leading to concerns regarding time efficiency, yet a reluctance to compromise on accuracy. This study evaluates selected fitting algorithms in terms of the precision/bias of fitted parameters and run‐time. As a standard for comparison, biological parameters were fixed at typical values (K1=0.2, k2=0.25, k3=0.05, Kᵢ=0.3334). Tissue time‐activity curves were generated using the model equation and by incorporating activity/time dependent Gaussian noise. The following algorithms were then applied: Genetic, Conjugate Gradient (CG), Gradient Descent (GD), Simulated Annealing (SA), Levenberg‐Marquardt (LMQ), Gauss‐Newton (GN), and Limited‐BFGS (L‐BFGS). Non‐iterative, problem specific approaches were also considered: Patlak analysis (K? only), and Blomqvist linearization (BL). Parameter accuracy and precision were quantified with relative errors (REs). Results: At typical noise levels, maximal REs were >60% (resulting in |Î′K?|>10%) for GD and SA, <20% for BL (|Î′K?|<6%) and Patlak (|Î′K?|<4%), and <15% (|Î′K?|<2%) with other methods. On average, the highest fidelity parameter estimates (rate‐constant REs<2.5%, |Î′K?|<0.15%) were attained with L‐BFGS, GN, and LMQ. In contrast, the non‐iterative methods provided poorer estimates (RE between 1–7%), though run‐time (<0.05 ms) was 100‐1000x less. Using BL results as initial estimates for L‐ BFGS (BL+L‐BFGS), GN (BL+GN) or LMQ (BL+LMQ) yielded rapid convergence (<0.1 ms), while maintaining superiority with respect to parameter bias. Conclusions: Poor choice of a fitting algorithm may lead to significant errors in estimated kinetic parameters and/or unpractically high computation times. Thus, our recommendation is the use of BL+L‐BFGS, BL+GN or BL+LMQ, which is contrary to the gold standards, Patlak and LMQ (alone), commonly seen in the literature.