High Temporal Resolution Total-Body Dynamic PET Imaging Based on Pixel-Level Time-Activity Curve Correction

Zixiang Chen, Yaping Wu, Na Zhang, Tao Sun, Yu Shen, Hairong Zheng, Dong Liang, Meiyun Wang, Zhanli HuIEEE Trans. Biomedical Engineering 2022 · DOI

TL;DR

This paper introduces "Pitch-In," a novel method for denoising dynamic PET (dPET) images with high temporal resolution. By focusing on pixel-level time-activity curve (TAC) correction using third-order Hermite interpolation, Pitch-In effectively reduces noise in short-frame dPET data without sacrificing image fidelity or structure, offering a promising solution for clinical applications.

What Problem Does This Paper Address?

Dynamic Positron Emission Tomography (dPET) is crucial for tracking radiotracer metabolism in real-time, aiding in cancer diagnosis and staging. However, achieving high temporal resolution (HTR) in dPET, which requires reconstructing images from very short time frames, leads to a significant trade-off: increased image noise and reduced signal-to-noise ratio (SNR). This bottleneck limits the clinical utility of HTR dPET, especially for capturing rapid metabolic changes. Existing methods often struggle to balance noise reduction with the preservation of image spatial resolution and temporal fidelity, leading to either over-smoothing or insufficient noise suppression. This paper aims to address this by developing a method that can effectively denoise dPET images with short frame durations while maintaining crucial image quality aspects.

Methods

The core of the proposed "Pitch-In" method lies in its pixel-level time-activity curve (TAC) correction strategy, which operates in the temporal domain rather than the spatial domain. This approach is based on Third-order Hermite Interpolation (THI), a technique that uses a wider range of reference points to estimate a target point, thus better preserving the original curve's tendency.

The Pitch-In method employs an iterative process that alternates between two key operations:

  1. THI Denoising Subloop: This subloop involves repeatedly applying a TAC re-estimation process using THI. For each pixel's TAC, a subset of data points is randomly selected as references, and the remaining points are re-estimated using THI. This process is repeated multiple times (epochs) to progressively denoise the TAC. A relaxation parameter ($\eta$) is used to update the TAC, balancing the denoised estimate with the previous iteration's result.
  2. TAC Fidelity Operation: To prevent over-smoothing introduced by the denoising process, a fidelity step is incorporated. This operation ensures that the denoised TACs remain consistent with the original noisy TACs. It defines "up-fidelity" and "bottom-fidelity" groups based on whether the denoised TAC values are lower or higher than the original TAC. Separate fidelity operations are then applied to these groups using predetermined relaxation parameters ($\beta_u$ and $\beta_b$) to correct deviations and maintain data fidelity.

These two operations are alternated within a main loop, allowing for an optimal trade-off between TAC smoothing and fidelity. The method is designed to be applied directly to the current imaging data, avoiding the need for training data or complex model generalization.

Datasets

The proposed Pitch-In method was evaluated using two types of datasets:

  1. Simulative dPET Data: Generated from an XCAT thoracic digital phantom. This phantom included nine simulated human tissues (aorta, bone, marrow, myocardium, ventricle, lung, liver, tumor, and background). Tissue-specific ground truth TACs were assigned, representing two time-dividing protocols:
  2. Low Temporal Resolution (LTR): 30 frames over one hour.
  3. High Temporal Resolution (HTR): 60 frames over one hour. The simulative images were reconstructed using the traditional expectation-maximization (EM) method.

  4. Clinical Total-Body dPET Data: Collected from a patient with lung adenocarcinoma using the uEXPLORER PET/CT system. The dynamic scan lasted for approximately one hour. Data were reconstructed using two protocols:

  5. HTR dPET: 84 frames.
  6. LTR dPET: 30 frames. A thoracic slice with a visible tumor lesion was used for the experiment.

Key Insights and Ideas

  • Temporal Domain Focus: The most significant insight is the shift from spatial denoising to temporal domain processing of TACs. This allows for the exploitation of dynamic information without the spatial blurring often associated with spatial denoising methods.
  • Balancing Smoothing and Fidelity: The iterative alternation between THI denoising and TAC fidelity operations is crucial. This ensures that noise is reduced effectively without distorting the underlying metabolic information or losing important structural details.
  • Model-Free Approach: The method's ability to work directly with current imaging data, without requiring prior training or complex model generalization, makes it highly practical for clinical applications.
  • Intensity-Independent Denoising: The Pitch-In method's denoising effect is not sensitive to the intensity of the original image signal, leading to more uniform denoising across tissues with varying radiotracer concentrations. This prevents unnatural over-smoothing of low-metabolism tissues.

Key Results

The Pitch-In method demonstrated superior performance in denoising HTR dPET images compared to several state-of-the-art methods (HYPR, STEM, NLM-ST, KIBM5D).

  • Simulative Experiments:
    • Visually, Pitch-In showed significantly better denoising than HYPR and STEM.
  • NLM-ST and KIBM5D smoothed background tissues more than hypermetabolic ones, while Pitch-In achieved spatial equilibrium.
  • Quantitative evaluation using PSNR and SSIM (Figure 4) showed Pitch-In to be superior in noise reduction and structure restoration across the whole image space and in specific ROIs (liver, ventricle, heart).
  • CNR values (Figure 5) indicated that Pitch-In achieved high contrast for most frames, especially for hypermetabolic tissues, and provided reliable ROI intensity distributions.
  • Clinical Image Experiments:
    • Pitch-In effectively eliminated image noise from noisy HTR images, outperforming other methods and even the LTR images.
  • Tissue structure was better restored, as evidenced by the clear aorta (marked by red arrows).
  • Background tissues like the lung were denoised without significant blurring.
  • CNR values for the aorta and tumor lesion (Figure 7) showed Pitch-In providing the best and most stable contrast.
  • Ki parametric images (Figure 8) indicated that Pitch-In did not seriously skew dynamic characteristics, with NLM-ST showing some blurring.

Limitations and Future Work

  • Potential for Flattening TACs: The authors acknowledge that if parameters are not well-adjusted, the method might flatten TACs with large net second-order derivatives (e.g., aorta, liver).
  • Future Work: To address this, they suggest developing an appropriate global cost function that combines TAC noise and distortion to derive a more effective and stable iterative TAC correction method.
  • Parameter Selection: While the method is robust, careful parameter selection is still necessary for optimal performance. The authors provide guidance on parameter selection and suggest that some parameters (like $k_p$, $n_r$, $E_t$) can be fixed, while others ($\eta$, $\beta_u$, $\beta_b$) play pivotal roles and require more careful tuning.
  • Clinical Feasibility: The paper highlights the method's high feasibility for clinical application due to its model-free nature and reliance on current data, contrasting it favorably with learning-based methods that may face generalization issues.

Summary updated

BibTeX
@article{chen2022high,
  author = {Zixiang Chen and Yaping Wu and Na Zhang and Tao Sun and Yu Shen and Hairong Zheng and Dong Liang and Meiyun Wang and Zhanli Hu},
  title = {{High Temporal Resolution Total-Body Dynamic PET Imaging Based on Pixel-Level Time-Activity Curve Correction}},
  journal = {IEEE Trans. Biomedical Engineering},
  year = {2022},
  doi = {10.1109/tbme.2022.3176097},
}
Figure