Ensemble Learning (EL) Independent Component Analysis (ICA) Approach to Derive Blood Input Function from FDG-PET Images in Small Animal

To extract the blood time-activity curves (TACs) from the PET image of a mouse heart is very difficult due to the limited spatial resolution of the PET system, small size of the heart, partial volume effects and cardiac motion. Ensemble learning-independent component analysis (EL-ICA), a recently developed Bayesian method, has been implemented to extract clear TACs from the PET images and also been proved to be a useful method for image segmentation. The advantage of EL-ICA is it decomposes the images into different independent components while imposing strong nonnegativity constraints, which can maintain the independence and nonnegativity of the component images and TACs simultaneously. A down-sampled, segmented CT data set has been used to generate simulated PET data to best represent the structure of a real cardiac image. From the results of the simulation, we can show that EL-ICA was able to extract the TACs of the sample data. We have also applied EL-ICA to FDG images in mice. In this study, we show that myocardium and blood pool components can be separated successfully by EL-ICA, and the according TACs obtained. The EL-ICA method can be used to extract the arterial input function directly from the dynamic PET images to avoid the need for multiple blood sampling of the small animal.