Abstract
3246
Introduction: PET kinetic modelling methods were developed several decades ago but and have been shown to contribute information relevant to disease diagnosis, prognosis and treatment planning. Yet they are not used widely in the clinic. Estimation of kinetic parameters requires dynamic PET (dPET) images, typically acquired over a period of 60 minutes. Shortened dPET protocols may facilitate clinical adoption of kinetic modelling. In 18F-FDG PET the time activity curve (TAC) generated for a volume of interest (VOI) or a voxel is used to estimate tissue specific parameters using the two tissue compartment model, namely the influx rate (K1, ml/g/min), rate constants (k2, k3, 1/min) and net influx rate (Ki=K1k3/(k2+k3), ml/g/min). Studies have investigated how well kinetic parameters can be fitted in shortened acquisition windows using the conventional non-linear least squares (NLSS) and machine learning based approaches. However, it remains unclear which parts of the TAC can be omitted to still achieve accurate estimates of kinetic parameters. Using supervised machine learning methods, we investigated the minimum TAC sampling requirements for accurate kinetic parameter estimation.
Methods: An in-house simulation environment in MATLAB® was created to produce synthetic data for 18F-FDG PET. Four thousand TACs were generated based on the two tissue compartment model with irreversible binding. Individual dPET TACs were simulated for 60 min (4 × 30s, 8 × 60s, 10 × 120s and 6 × 300s frames). Parameters K1, k2 and k3 were selected based on values from previously reported ranges. The blood fraction in tissue, Vb, was randomly chosen in [0.02, 0.05]. Noise was introduced to each TAC using a previously described approach. Testing was performed on a digital dPET brain phantom with realistic grey-white matter kinetic parameters. Neighbourhood Component Analysis (NCA) was used to establish the sensitivity of kinetic parameter estimation to data collection time. Different TAC durations were also evaluated by systematically increasing time intervals up to 60 min. Gaussian Process Regression (GPR) appended by an autoencoder was employed to build the kinetic modelling framework (Figure 1A). Individual TACs and the arterial input function formed the feature vectors, and corresponding kinetic parameters, K1, k2, k3 and Ki were the labelled data. Goodness of fit was assessed using the R2 value, and error in the kinetic parameter estimates was recorded.
Results: Figure 1B summarises the findings for shortened acquisition times based on results provided in Figures 1C and 1D. Figure 1C illustrates the NCA output charts for each of the kinetic parameters based on a 60 min data collection schedule. The chart shows that the highest index feature weights for K1 and k2 occur at early time points, i.e., 0 to 5 min, and k3 estimation requires the first 8 minutes and last 15 minutes of a 60 min imaging window. Our NCA results confirmed that the initial imaging window, 0 to 8 min, contained critical information for estimating all kinetic parameters. While the final minutes of the imaging window is essential for estimating k3, and less critical for k2 and Ki, and not necessary for K1. Figure 1D provides the R2 results for estimated kinetic parameters when the TAC duration is increased from 2 min to 60 min. These results confirm that K1 can be estimated using a short window, while k3 requires longer acquisitions. Figure 2 depicts parametric maps obtained for the brain phantom across the nominated protocols defined in Figure 1B. The results in Figures 1C, 1D and 2 confirm that the double imaging protocol [0-8 min and 55-60 min] and a single imaging window protocol [0-24 min] provide similar accuracy in estimating the kinetic parameters (R2>~0.90) as the standard dPET protocol.
Conclusions: Results provided suggest that the proposed machine learning kinetic modelling framework, including GPRs and autoencoders, provides robust kinetic parameter estimates in shortened imaging windows identified using NCA.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.