Quantitative dynamic cardiac 82Rb PET using generalized factor and compartment analyses.

Georges El Fakhri, Arkadiusz Sitek, Bastien Guérin, Marie Foley Kijewski, Marcelo F. Di Carli, Steven C. MoorePubMed 2005

TL;DR

This paper introduces a novel approach combining Generalized Factor Analysis of Dynamic Sequences (GFADS) with a 2-compartment kinetic model to quantitatively assess myocardial perfusion using dynamic 82Rb PET scans. This method overcomes limitations of traditional techniques by automatically estimating the input function and providing voxel-wise parametric maps, achieving high accuracy (<9%) without arterial blood sampling.

What Problem Does This Paper Address?

Assessing myocardial perfusion with 82Rb PET is valuable for diagnosing coronary artery disease, especially in clinics lacking cyclotrons. However, accurate quantification is challenging due to two main issues: 1) estimating a precise input function from noisy PET data without invasive arterial blood sampling, and 2) achieving absolute quantification of myocardial perfusion. Existing methods often rely on manual region-of-interest (ROI) definitions, which are prone to noise and inter-observer variability, and can struggle with accurate modeling of tracer kinetics in the myocardium.

Methods

The authors propose a two-pronged approach:

  • Generalized Factor Analysis of Dynamic Sequences (GFADS): This technique is used to automatically extract the left (LV) and right (RV) ventricular time-activity curves (TACs) directly from the dynamic PET data. GFADS is a variation of factor analysis that incorporates a penalty term to minimize spatial overlap between the estimated factor images, thereby improving the uniqueness and robustness of the solution. This avoids the need for manual ROI drawing.
  • 2-Compartment Kinetic Modeling: Once the LV and RV TACs (serving as input functions) are estimated by GFADS, a 2-compartment kinetic model is applied on a voxel-by-voxel basis. This model estimates key physiological parameters:
  • k₁: Myocardial tissue extraction rate (inflow).
  • k₂: Myocardial tissue egress rate (outflow).
  • fᵥ: LV contribution to the voxel's activity.
  • rᵥ: RV contribution to the voxel's activity.

The kinetic model accounts for both the trapped rubidium space (myocardium) and the free circulating space (blood). To handle noise in voxel-wise TACs, an orthogonal grouping method is employed, where voxels with similar dynamic behavior are grouped, and the average TAC of the group is used for parameter estimation.

Datasets

  • Monte Carlo Simulations: Thirty realistic dynamic 82Rb PET studies were generated using a 2D PET Monte Carlo simulation (SimSET) based on the anthropomorphic Zubal torso phantom. These simulations incorporated realistic LV, RV, myocardium, and other organ activity distributions, as well as known kinetic parameters (k₁, k₂) and input functions. This dataset was used to evaluate the performance of the GFADS and kinetic modeling approach.
  • Patient Studies: Data from 13 patients were analyzed. Five patients had evidence of obstructive coronary artery disease (CAD), and eight did not. Rest and stress dynamic protocols were used. This dataset was used to demonstrate the clinical feasibility and robustness of the method.

Key Insights and Ideas

  • Automated Input Function Estimation: GFADS effectively bypasses the need for arterial blood sampling or manual ROI definition for input function estimation, a significant practical advantage.
  • Robustness to Noise: The GFADS approach, combined with orthogonal grouping, demonstrated superior robustness to noise compared to traditional VOI-based methods, yielding more accurate and stable estimates.
  • Voxel-wise Quantification: The ability to perform kinetic modeling on a voxel-by-voxel basis allows for detailed mapping of myocardial perfusion parameters, potentially identifying regional differences.
  • Improved Accuracy: The proposed method achieved significantly lower errors in estimating input functions and myocardial TACs compared to VOI-based methods, particularly in scenarios with spillover.
  • Physiological Relevance: The estimated kinetic parameters (k₁, k₂, fᵥ, rᵥ) were found to correlate well with clinical status and catheterization results, indicating physiological relevance.

Key Results

  • Monte Carlo Simulations:
    • GFADS yielded estimates of LV, RV, and myocardial factors with average errors of -4.2% ± 6.3%, 3.5% ± 4.3%, and 2.0% ± 5.5%, respectively.
  • The 2-compartment kinetic analysis yielded parametric maps with average errors of <9% for k₁, k₂, fᵥ, and rᵥ. Specifically, k₁ error was 6.8% ± 3.6%.
  • GFADS estimates were significantly more accurate and robust to noise than VOI-based estimates (P < 0.01), with VOI errors for LV, RV, and myocardial factors being much higher (e.g., 30.7% ± 8.3% for LV).
  • Patient Studies:
    • The method provided robust estimates of k₁, k₂, fᵥ, and rᵥ that correlated well with patient status.
  • Coronary Flow Reserve (CFR) values were systematically higher in subjects without CAD (2.1 ± 0.3) compared to those with known CAD (1.4 ± 0.1).
  • The overall accuracy for quantitative dynamic 82Rb PET without arterial sampling was demonstrated to be less than 9%.

Limitations and Future Work

  • Number of Factors (P): While P=3 was found to be sufficient for robust estimates in simulations and patient studies, the authors plan to investigate the potential of using a greater number of factors to describe more complex dynamic activity distributions.
  • Cluster Size (G): The choice of cluster size G was found not to be critical, with G=30 and G=100 working well. However, the authors suggest further investigation into optimal cluster size.
  • Respiratory Motion and Cardiac Motion: The authors acknowledge that respiratory motion and cardiac motion can affect tissue count recovery and potentially lead to similar time curves representing different tracer kinetics. They suggest using spatial constraints in the orthogonal grouping analysis as a potential solution.
  • Dead Time: The effect of dead time, which is important in high-count-rate scenarios like 82Rb PET, was not modeled in the Monte Carlo simulations but is an area of ongoing work.
  • Flow Dependency of Extraction: While the current model assumes a linear relationship between flow and extraction, the authors note that this assumption may not hold at very high flows. They suggest that this flow dependency could be incorporated as a postprocessing step or a priori knowledge.
  • Regional Discrimination: While GFADS provides independent factor images for LV, RV, and myocardium, the current kinetic model uses a single factor for myocardial TAC. The authors suggest that combining GFADS with voxel-wise kinetic analysis allows for discrimination between normal and diseased myocardial blood flow, but further work might be needed to fully characterize regional differences within the myocardium.

Summary updated

Abstract

Quantitative dynamic (82)Rb PET using generalized factor analysis of dynamic sequences and compartmental modeling yields estimates of parameters of absolute myocardial perfusion and kinetics with errors of <9%.
BibTeX
@inproceedings{fakhri2005quantitative,
  author = {Georges El Fakhri and Arkadiusz Sitek and Bastien Guérin and Marie Foley Kijewski and Marcelo F. Di Carli and Steven C. Moore},
  title = {{Quantitative dynamic cardiac 82Rb PET using generalized factor and compartment analyses.}},
  booktitle = {PubMed},
  year = {2005},
}
Figure