Effects of regularisation priors on dynamic PET Data

Liliana Caldeira, J. Scheins, Nuno Silva, Michaela Gaens, N. Jon ShahEJNMMI 2014 · DOI

TL;DR

This paper investigates how regularization priors can improve the quality of dynamic Positron Emission Tomography (PET) images, which are often noisy due to low photon counts in individual frames. By applying a Median Root Prior (MRP) during reconstruction, the researchers significantly reduced image noise and smoothed parametric images, leading to better visualization and potentially more accurate quantification.

What Problem Does This Paper Address?

Dynamic PET imaging captures how a radioactive tracer moves through the body over time, providing valuable functional information. However, a key challenge is that each individual time frame in a dynamic PET scan contains a limited number of detected photons (counts). This low count statistics leads to significant image noise, making it difficult to accurately assess tracer uptake and derive meaningful quantitative parameters, especially in early frames. This noise can obscure subtle changes and affect the reliability of downstream analyses, such as generating parametric maps that represent physiological processes.

Methods

The study employed a simulated dynamic PET dataset. The core of their method involved comparing two reconstruction techniques:

  • Ordinary Poisson Ordered Subset Expectation Maximization (OP-OSEM): This is a standard iterative reconstruction algorithm for PET data.
  • MRP OP-OSEM: This is an enhanced version of OP-OSEM that incorporates a Median Root Prior (MRP). The MRP is a type of regularization that uses a 3x3x3 neighborhood to smooth the image and a "Bayes parameter" to control the strength of this smoothing. The authors applied this prior to all frames of the dynamic PET data.

Parametric images, specifically Patlak parametric images, were generated using PMOD software to assess the impact of the reconstruction methods on quantitative measures.

Datasets

The study used a simulated dynamic PET dataset. This dataset was generated to mimic real-world [18F]-FDG PET data acquired on a Siemens 3T MR-BrainPET scanner. The simulation incorporated:

  • Realistic time-activity curves (TACs) for two tissue types (White Matter and Gray Matter) and the arterial input function, derived from actual volunteer data.
  • A total of 6x10^8 true counts.
  • Scatter and random fractions of 35% each.
  • 23 frames, representing a typical clinical protocol.

No new benchmark datasets were introduced; the focus was on evaluating reconstruction methods on a simulated, yet realistic, dataset.

Key Insights and Ideas

  • Noise Reduction is Significant, Especially in Early Frames: The most striking insight is the substantial reduction in image noise achieved by the MRP prior, particularly in frames with fewer counts. This directly addresses the core problem of low count statistics in dynamic PET.
  • Parametric Images Benefit from Smoothing: The regularization not only cleans up the individual PET frames but also leads to smoother and potentially more interpretable parametric images (like Patlak maps). This suggests that the benefits extend beyond visual quality to quantitative analysis.
  • Count-Dependent Behavior of Regularization: The study observed that the relative performance of MRP OP-OSEM versus standard OP-OSEM in terms of quantification (measured by RCs - likely Recovery Coefficients or similar metrics) shifted depending on the frame's count statistics. MRP OP-OSEM showed higher RCs in low-count frames, while OP-OSEM was slightly higher in high-count frames. This highlights that the effectiveness and impact of regularization can be dynamic and count-dependent.

Key Results

  • Image Noise Reduction: The MRP OP-OSEM method achieved image noise reduction ranging from 50% in low-count frames to 10% in high-count frames (e.g., Frame 24).
  • Parametric Image Quality: Patlak parametric images generated with MRP OP-OSEM appeared smoother compared to those from standard OP-OSEM.
  • Quantification: Both reconstruction methods showed similar overall quantification. However, MRP OP-OSEM yielded higher RCs in low-count frames (up to 9%), while OP-OSEM showed slightly higher RCs in high-count frames (up to 9%).

Limitations and Future Work

  • Simulation-Based Study: The primary limitation is that the study was conducted using simulated data. While the simulation was designed to be realistic, it may not perfectly capture all the complexities and artifacts present in real-world PET data.
  • Specific Prior: The study focused on the Median Root Prior (MRP). Other types of regularization priors might yield different results or offer alternative trade-offs between noise reduction and image detail.
  • Bayes Parameter Optimization: The Bayes parameter for the MRP was set to a fixed value (0.1). Further investigation into optimizing this parameter, potentially on a frame-by-frame basis or in a count-dependent manner, could be beneficial.
  • Future Work: The authors suggest that this study demonstrates the potential of prior regularization for improving dynamic PET image quality. Future work could involve validating these findings on real patient data, exploring other regularization techniques, and investigating the impact of these improvements on specific clinical applications and diagnostic accuracy.

Summary updated

Abstract

Dynamic PET provides temporal information about tracer uptake. However, each PET frame has usually low statistics, resulting in noisy images. The goal is to study effects of prior regularisation on dynamic PET data. Quantification and noise in image-domain and time-domain as well as impact on parametric images is assessed. Dynamic PET data for the Siemens 3T MR-BrainPET was simulated with time-activity curves (TAC) of [1] F-FDG obtained analytically using realistic values for the kinetic model (Table ​(Table1).1). The total number of true counts was 6x108 and scatter and random fractions were both 35%. The data consists of 23 frames as applied in our clinical protocol. For reconstruction, the Ordinary Poisson Ordered Subset Expectation Maximisation (OP-OSEM) method was used in PRESTO [1], which allows to use several 3D priors [2]. The Median Root Prior (MRP) used a 3x3x3 neighbourhood and a Bayes parameter of 0.1 for all frames. Patlak parametric images were calculated using PMOD software. Table 1 Kinetic values used for simulation of two tissues: White-Matter (WM) and Gray-Matter (GM), extracted from real acquired volunteer data. Figure ​Figure11 shows that the MRP OP-OSEM reduces image noise in WM, GM and AIF: from 50% reduction in low-count frames to 10% reduction in high-count frames (e.g. Frame 24). Furthermore, Patlak parametric images also look smoother for MRP OP-OSEM than for OP-OSEM. Figure 1 a) Reconstructed images without MRP corresponding to Frame 14 (2.77x106 counts); b) Reconstructed images with MRP corresponding to Frame 14 (2.77x106 counts); c) Reconstructed images without MRP corresponding to Frame 24 (119x106 counts); d) Reconstructed ... Figure ​Figure22 shows similar quantification for both reconstruction methods (with similar RCs). OP-OSEM presents higher RCs than MRP OP-OSEM in high-count frames (up to 9%), while in low-count frames MRP OP-OSEM presents higher RCs (up to 9%). This is probably due to the fact that the Bayes parameter is count-dependent. Figure 2 TAC using OP-OSEM reconstruction without and with MRP for two different regions: a) GM and b) AIF. This study shows improvement on PET image quality in terms of noise (up to 50% reduction) as well as in parametric images when using prior regularisation in dynamic PET data. Both OP-OSEM and MRP OP-OSEM show similar quantification, with higher RCs for MRP OP-OSEM in low-count frames.
BibTeX
@inproceedings{caldeira2014effects,
  author = {Liliana Caldeira and J. Scheins and Nuno Silva and Michaela Gaens and N. Jon Shah},
  title = {{Effects of regularisation priors on dynamic PET Data}},
  booktitle = {EJNMMI},
  year = {2014},
  doi = {10.1186/2197-7364-1-s1-a46},
}
Figure