Evaluation of factor analysis accuracy for myocardial perfusion in PET studies

Abstract

Factor Analysis of Medical Image Sequences (FAMIS) estimates kinetics ( factors) and corresponding spatial regions (factor images) from dynamic studies, taking into account statistical noise and spillover effects. Factor images obtained from 150-water clinical cardiac PET studies are less noisy than conventional subtraction images, and factors match physiologic kinetics. The aim of this work was to study FAMIS accuracy and precision depending on the application context. They were evaluated through kinetics parameters quantification. Numerical simulations and phantom experiments were carried out using a typical left ventricular pattern. This object was simulated in 2D with three noise levels and two kinds of kinetics : mono-exponentials which correspond to natural tracer decay and tissue perfusion kinetics obtained with a realistic vascular input function. Monoexponentials association was adapted to phantom experiments while perfusion kinetics standed for clinical cardiac studies. In both phantom experiments and simulations, the inner chamber was filled with 150-water and the myocardial space with Carbon-1 1. The different noise levels which were studied corresponded to ideal, normal and low quality scans. Using the factors estimated by FAMIS, decay constants and an index of flow (kl) were estimated by fitting or modelling. Relative bias to the true value and standard deviation were then estimated, and spatial correlation between factor images and original spatial pattern was computed. Factor images spatial correlation was very good, despite of large overlapping pattern. Oxygen-15 decay constant was assessed from factor with a small relative bias, for all noise levels and trixel sizes. However, Carbon-1 1 extraction was very sensitive to both noise and spillover in phantom and simulations. A reasonable bias was only achieved by including a spillover term which accounted for an overcorrection. On the contrary, factors associated to perfusion were well extracted and k l parameter was recovered with a low relative bias (r,b. < 6%), except for the higher noise level. It was already shown that FAMIS performences depend on the overlap of the spatial structures.This study demonstrates that factor analysis without a priori information performances depend on kinetics shape. Moreover, in the context of cardiac 150-water perfusion studies, FAMIS should provide accurate quantification.

Segmentation of dynamic PET images using cluster analysis

“…Factor analysis (FA) can be thought of as a generalisation of PCA as it produces factors closer to the true underlying tissue TACs. It has been used to extract the time course of blood activity in the left ventricle of the heart from PET images [G] and to analyse myocardial perfusion in PET studies [7].…”