A modeling-based factor extraction method for determining spatial heterogeneity of Ga-68 EDTA kinetics in brain tumors

Abstract

The ROI method applied to Ga-68 EDTA PET dynamic study data for the quantitative determination of brain tumor BBB permeability assumes that the tumor is homogeneous in terms of Ga-68 EDTA kinetics, even though it is known that brain tumors are highly heterogeneous in structure. The relatively high image noise of Ga-68 PET studies have prevented the examination of Ga-68 EDTA kinetics by nonlinear regression on a single pixel basis. In this study, we have developed an efficient and effective method to separate brain tumor tissue into sub-regions with different Ga-68 EDTA kinetics on a pixel-by-pixel basis. Computer simulation and ten Ga-68 EDTA PET patient studies were used to evaluate the performance of the new method. During a PET dynamic study (total 64 min), 20-25 arterial samples were taken for the input function. Whole-tumor ROIs were defined on T1-weighted MRI images and then copied to the registered PET dynamic images to measure whole-tumor time activity curves. The method uses a two-compartment model to extract three component factors (vascular component, fast and slow component factors) from whole-tumor kinetics by model fitting. The kinetics in each pixel were expressed as a linear combination of the three factors. The three coefficients in the expression can be estimated by the linear least-square method and produce three factor images corresponding, respectively, to the permeability of the fast and the slow component factors and the plasma volume. Whole-tumor regions were separated into two regions-one with mainly fast kinetics that was evident in the fast factor images and one with slow kinetics that was evident in slow factor images. The two regions have markedly different uptake (0.036/spl plusmn/0.015 ml/min/g and 0.009/spl plusmn/0.006 ml/min/g for fast and slow kinetic sub-regions, respectively) and clearance rates (0.22/spl plusmn/0.15 /min and 0.023/spl plusmn/0.021 /min for fast and slow sub-regions, respectively). The overlap of the two resulting sub-regions is small (3.5/spl plusmn/1.7% of the two regions). Computer simulation and patient studies show that the method is robust for a wide range of noise levels. This method has combined the advantages of statistical factor analysis and the modeling approach.