A Bayesian spatial temporal mixtures approach to kinetic parametric images in dynamic Positron Emission Tomography

Wanchuang Zhu, Jinsong Ouyang, Yothin Rakvongthai, Nicolas J. Guehl, Dustin Wooten, Georges El Fakhri, Marc D. Normandin, Yanan Fan

Abstract

We present a fully Bayesian statistical approach to the problem of compartmental modelling in the context of Positron Emission Tomography. We cluster homogeneous region of interest and perform kinetic parameter estimation simultaneously. A mixture modelling approach is adopted, incorporating both spatial and temporal information based on reconstructed dynamic PET image. Our modelling approach is flexible, and provides uncertainty estimates for the estimated kinetic parameters. Crucially, the proposed method allows us to determine the unknown number of clusters, which has a great impact on resulting estimated kinetic parameters. We demonstrate our method on simulated dynamic Myocardial PET data, and show that our method is superior to standard curve-fitting approach.

Introduction

Dynamic Positron Emission Tomography (PET) imaging provides a temporal sequence of 3D images, essentially a 4D "movie" tracking the distribution of a radioactive tracer within tissue. By fitting these time-activity curves (TACs) to compartmental models, researchers can estimate critical physiological parameters such as perfusion ($K_1$), which represents the rate of blood delivery to an organ.

However, generating high-quality kinetic parametric images is challenging due to the high noise levels inherent in PET data. Voxel-wise estimation, where parameters are calculated for each pixel independently, often results in extremely noisy and "grainy" images. Conversely, traditional Region of Interest (ROI) estimation averages large areas to reduce noise but sacrifices spatial resolution and risks blurring distinct tissue types. This paper introduces a Bayesian Spatial Mixture Model (SMM) designed to find an optimal balance by simultaneously clustering voxels and estimating parameters.

Methodology: The Bayesian Spatial Mixture Model

The proposed SMM framework addresses the noise-resolution trade-off through several key innovations:

  1. Simultaneous Clustering and Estimation: Instead of a two-step process, the model uses Markov chain Monte Carlo (MCMC) to identify voxel groups and estimate their kinetic parameters concurrently.
  2. Spatial Priors (Potts Model): To reflect biological reality, a Markov Random Field (specifically a Potts model) is employed. This encourages neighboring voxels to belong to the same cluster, effectively "borrowing" information from adjacent pixels to suppress noise.
  3. Automated Model Selection: The number of clusters ($G$) is not fixed. The model utilizes the Bayesian Information Criterion (BIC) to objectively determine the optimal number of clusters supported by the data.
  4. Noise Handling: A dedicated "noise cluster" is included to capture background signals or voxels outside the target organ, preventing them from biasing the physiological results.

Experimental Evaluation

The SMM was evaluated using both simulated and real-world data.

Simulated Cardiac PET Data

A 13-minute dynamic PET scan was simulated using the NCAT torso phantom. This provided a "ground truth" where the exact kinetic parameters for various heart segments were known, allowing for precise bias and error measurements.

Time-activity curves (TACs) for the input function, normal tissue, and defect tissue in the simulated cardiac PET data.

The simulation included various anatomical segments, each with specific kinetic properties. The model's ability to recover these parameters was tested against standard methods.

True kinetic parameter values ($K_1$ and $k_2$) for different anatomical segments used in the simulated phantom study.

In-Vivo Pig Study

The algorithm was also applied to a real-world dataset from a 40kg pig scanned with the tracer $^{18}$F-flurpiridaz. This test assessed the model's performance in a biological setting where ground truth is unknown but physiological plausibility (e.g., smooth perfusion maps) can be assessed.

Key Results

The SMM was compared against Standard Curve Fitting (SCF) and Spatial K-means (SKMS):

  • Reduced Variability: In simulations, the SMM achieved a median standard deviation of 0 for $K_1$ estimates across voxels, a dramatic improvement over SCF (0.13) and SKMS (0.16).
  • Improved Accuracy: In diseased heart regions, SMM showed a lower median mean squared bias (0.04) compared to both SCF and SKMS (0.06).
  • Physiological Plausibility: In the pig study, SMM produced smoother parametric maps that avoided the unrealistic "truncated" values often seen with standard curve-fitting methods.

Conclusion and Future Work

The Bayesian Spatial Mixture Model provides a robust framework for generating kinetic parametric images. By integrating spatial information and automated clustering, it significantly reduces noise while maintaining physiological accuracy. The primary trade-off is computational time, with the SMM requiring significantly longer processing than standard methods. Future work will focus on optimizing these algorithms and relaxing the homogeneity assumption to allow for natural biological gradients within clusters.

Summary updated

Abstract

We present a fully Bayesian statistical approach to the problem of compartmental modelling in the context of Positron Emission Tomography. We cluster homogeneous region of interest and perform kinetic parameter estimation simultaneously. A mixture modelling approach is adopted, incorporating both spatial and temporal information based on reconstructed dynamic PET image. Our modelling approach is flexible, and provides uncertainty estimates for the estimated kinetic parameters. Crucially, the proposed method allows us to determine the unknown number of clusters, which has a great impact on resulting estimated kinetic parameters. We demonstrate our method on simulated dynamic Myocardial PET data, and show that our method is superior to standard curve-fitting approach.
BibTeX
@inproceedings{zhu2016bayesian,
  author = {Wanchuang Zhu and Jinsong Ouyang and Yothin Rakvongthai and Nicolas J. Guehl and Dustin Wooten and Georges El Fakhri and Marc D. Normandin and Yanan Fan},
  title = {{A Bayesian spatial temporal mixtures approach to kinetic parametric images in dynamic Positron Emission Tomography}},
  booktitle = {Medical},
  year = {2016},
  doi = {10.1118/1.4941010},
}
Figure