Abstract
1932
Objectives We developed a new algorithm of iterative factor analysis, which enabled to estimate non-invasive measure of the input function in dynamic image data. Factor analysis has been tried to estimate input function and tissue component curve in the dynamic image data, however, it is hard to extract more than two independent factors robustly. Then we developed the iterative two-factor analysis to estimate the input function robustly. To evaluate the algorithm, dynamic 18F-FDG brain PET data were adopted, cerebral metabolic rate of glucose (CMRGlc) was calculated using input functions both from sampled blood data and from estimated curve with iterative factor analysis.
Methods Ten patients (36.8±13.8 years old, male 5, female 5 ) with neurological diseases underwent 18F-FDG dynamic brain PET study for 60 minutes and serial blood sampling. CMRGlc were calculated on 100 ROIs in each patient's brain image with three compartment model analysis using time-activity plasma data. CMRGlc were recalculated on the same ROIs using the input function estimate d from the iterative two-factor analysis. The program of iterative two-factor analysis was developed by ourselves using Microsoft Visual C#. The first two-factor analysis was performed in the 32x16 pixels (5x2cm) rectangular ROI, which was set manually between the bilateral temporal bases, enclosing bilateral internal carotid arteries, cavernous sinuses and cerebral temporal lobes. The first factor analysis extracted the independent two factor curves and factor images which corresponded to the cerebral cortical component and blood component. The second two-factor analysis was performed only in the ROI data of the blood component image, the second factor analysis yielded the independent two factor curves and factor images which corresponded to the arterial blood component and cavernous sinus component. The third and the fourth two-factor analysis were performed only in the ROI data of the arterial blood component image, the third and fourth two-factor analysis extracted the independent two factor curves and factor images which corresponded to the arterial blood component in the internal carotid arteries and the surrounding venous blood component. Extracted arterial blood curve was multiplied by 1.1 to convert plasma curve data. The dynamic 18F-FDG brain PET data and estimated plasma curve from the second, third and fourth factor analysis were used to estimate CMRGlc on 100 ROIs in each patient's brain image with three compartment model analysis.
Results The factor curves and images which corresponded to the internal carotid artery by factor analysis with second, third and fourth iterations were estimated. They were compared with the sampled arterial blood curve, the curves from second and third iterations presented the mixed data of arterial blood, venous blood and surrounding tissue component, however, the fourth iteration yielded the carotid arterial curve which was almost same as the sampled arterial blood curve. Correlation coefficients between CMRGlc from sampled blood and from input curve by factor analysis with second, third and fourth iterations were 0.59±0.24, 0.74±0.17 and 0.92±0.03, respectively. The differences in these correlation coefficients were compared, factor analysis with fourth iteration yielded significantly high coefficient (P<0.001, ANOVA), We found that fourth iteration of factor analysis dedicated appropriate estimation of input function for dynamic 18F-FDG brain PET.
Conclusions Using iterative two-factor analysis, we established a method for estimating input function with 18F-FDG dynamic brain PET data. Two-factor analysis enabled to yield mathematically robust extraction of independent factor images and curves. And the iterative factor analysis enabled to yield the accurate factor image by taking multiple extracting steps. The iterative factor analysis will be useful method for various dynamic medical data.