TY - JOUR T1 - Reference Region Kinetic Analysis Method based on an extension of the Likelihood Estimation in Graphical Analysis (LEGA) JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 363 LP - 363 VL - 58 IS - supplement 1 AU - Jean-Dominique Gallezot AU - Richard Carson Y1 - 2017/05/01 UR - http://jnm.snmjournals.org/content/58/supplement_1/363.abstract N2 - 363Objectives: The Logan graphical analysis (GA) [1], and its non-invasive (reference region) extension (NIGA) [2] are methods to estimate the volume of distribution (VT) and binding potential (BPND) of reversible tracers, respectively. In theory, these two methods can be applied to any reversible tracer, without needing to choose a particular compartmental model configuration, and only requires estimating 2 (for GA) or 3 (for NIGA) parameters. However, noisy tissue data are used in the independent variable in GA and NIGA, which leads to noise-induced bias of the VT or BPND estimates [3], especially in regions with high binding or slow kinetics, or when using very noisy time-activity curves (TACs), such as from a single voxel. The multilinear analysis 1 (MA1) [4] and the multilinear reference tissue model (MRTM) [5] partially mitigate this issue for GA and NIGA, respectively, by using less noisy tissue data in the independent variables. The LEGA method fully mitigated this issue for GA, by not using any noisy data in the independent variable. In this study, we present a hybrid reference tissue model, HRTM, which is an extension of LEGA for NIGA.Methods: New equations for LEGA and its non-invasive extension were obtained using Laplace transforms to solve the implicit differential equation in GA and NIGA. The HRTM equation has the form of a simplified reference tissue model (SRTM), only valid for time > t[asterisk], with an extra exponential term. Also at t=t[asterisk], the equation is the same as the MRTM equation. Hence, the method is denoted as a Hybrid of SRTM and MRTM. We compared the performance of NIGA, MRTM and HRTM using real human [11C]P943 data (n=25) and simulations. For simulations, curves were generated using a two-tissue compartment model, and Gaussian noise was added taking into account the shape of the noise-equivalent count curve from a typical study. The variance of the noise was scaled by a factor ranging from 0.01 (noise level of TACs from large regions of interest) to 22 (low dose single voxel TACs). One thousand noise realizations were performed.Results: On real human [11C]P943 regional TAC data, the agreement between MRTM and HRTM was very strong (the slope of the regression line of HRTM BPND estimates versus MRTM BPND estimates was 0.991, the intercept was 0.003 and the Pearson correlation coefficient was 1.000). Conversely, NIGA BPND estimates tended to be lower, i.e., underestimated as expected, especially in high binding regions (the slope of regression line of NIGA BPND estimates versus MRTM BPND estimates was 0.787, the intercept was 0.165 and the Pearson correlation coefficient was 0.928). On simulated data corresponding to the calcarine cortex (cortical region with the highest BPND for [11C]P943), the distribution of MRTM and HRTM BPND estimates were similar at low to mid noise levels (identical median, and 5%, 25%, 75% and 95% percentile up to noise level 0.48). At higher noise level, HRTM had lower noise, i.e. narrower 25-75% and 5%-95% percentile intervals than MRTM. As expected, NIGA BPND estimates were increasingly biased with increasing level of noise. Conversely, the median (mean, respectively) of BPND values were within 5% of the simulated value up to noise level 4.8 (1.0, respectively) for both HRTM and MRTM.Conclusion: The new non-invasive extension of the LEGA method, HRTM, has a similar or better performance as MRTM. An additional advantage of HRTM is that its fitted curves are not affected by noise in the tissue data, unlike fitted curves produced by MA1 or MRTM. Moreover, knowledge of the tissue curve is not needed in HRTM to compute model curves, thus HRTM is potentially well suited for incorporation into direct reconstruction algorithms, where additional noise reduction would be expected. Research Support: N/A ER -