Project Summary/Abstract
Dynamic PET myocardial perfusion imaging is emerging as a new modality to quantitatively measure
myocardial blood flow and flow reserve and has improved the diagnostic and prognostic assessments of
coronary artery disease (CAD) compared with other imaging techniques. However, flow quantification by the
standard three-dimensional (3D) approach faces a substantial noise challenge due to further dividing the noisy
tomographic data into shorter frames for image reconstruction before time activity curves are extracted for
kinetic analysis. Four-dimensional (4D) parametric image reconstruction incorporates kinetic modeling into
directly estimating parametric images from entire raw data. It has statistical advantages over the indirect 3D
method by accurately modeling noise in the projection space. However, direct parametric image estimation still
suffers from the inherent ill-posedness of the reconstruction problem, to which spatial regularization extensively
developed in static PET imaging has rarely been explored. Furthermore, 4D direct reconstruction has typically
been investigated at the concept implementation stage with very limited animal or patient study validation.
In this project, we propose to develop novel data-driven motion corrected direct 4D parametric image
reconstruction techniques integrating unsupervised deep learning-based regularization to reduce noise while
maintaining quantitative accuracy in dynamic 82Rb PET MP imaging. We will develop and validate the
techniques in terms of global and regional noise reduction using large animal models of ischemia and
myocardial infarction (MI), patients with suspected ischemic heart disease (IHD), and patients following MI
referred for assessment of viability. The performance of the proposed method will be compared with the
standard indirect 3D approach and with non-regularized and other regularized 4D methods. We hypothesize
that the proposed techniques will significantly reduce measurement uncertainty of 82Rb kinetics, leading to
improved test-retest repeatability, diagnostic accuracy in evaluating IHD, and reliability in accessing myocardial
viability with rest 82Rb MP imaging alone.
The proposed learning regularized direct 4D parametric reconstruction will be the first attempt to integrate
deep learning-based denoising to resolve the low signal to noise ratio challenge in dynamic PET imaging. The
advancements brought by this project will enable dynamic 82Rb imaging for 1) detection and risk stratification of
patients with obstructive CAD and ischemia with non-obstructive coronary arteries (INOCA), and 2)
identification of myocardial viability in patients with IHD referred for potential revascularization. In patients
referred for evaluation of myocardial viability, the proposed method will discriminate hibernating myocardium
from scar with a single rest 82Rb PET scan. The proposed efforts will revolutionize the evaluation and
management of patients with suspected or known CAD or those being considered for coronary
revascularization to increase clinical efficiency and reduce healthcare cost.