Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dual Prompting for Diverse Count-level PET Denoising

Xiaofeng Liu, Yongsong Huang, Thibault Marin, Samira Vafay Eslahi, Tiss Amal, Yanis Chemli, Keith Johnson, Georges El Fakhri, Jinsong Ouyang

TL;DR

The paper tackles the challenge of denoising low-count PET images across a continuous range of count levels. It introduces a dual-prompt framework combining a Specific Count Level Prior Prompt $\mathcal{P}_C = \omega_C(\delta)$ and a Learnable General PET Denoising Prompt $\mathcal{P}_G$, fused via cross-attention and injected into a 3D SwinUNet denoiser through skip connections. Key contributions include the explicit CLP mapping, the learnable GPD prompt, cross-attention fusion, and prompt injection strategy, all trained with an $L_1$ loss. Empirical results on $^{18}$F-MK6240 tau PET data show that the dual-prompt approach outperforms count-level conditioning and prior prompt methods, demonstrating improved MAE, MSE, PSNR, and 3D SSIM across 13–22% count levels and suggesting broader applicability to degradation-aware medical imaging tasks.

Abstract

The to-be-denoised positron emission tomography (PET) volumes are inherent with diverse count levels, which imposes challenges for a unified model to tackle varied cases. In this work, we resort to the recently flourished prompt learning to achieve generalizable PET denoising with different count levels. Specifically, we propose dual prompts to guide the PET denoising in a divide-and-conquer manner, i.e., an explicitly count-level prompt to provide the specific prior information and an implicitly general denoising prompt to encode the essential PET denoising knowledge. Then, a novel prompt fusion module is developed to unify the heterogeneous prompts, followed by a prompt-feature interaction module to inject prompts into the features. The prompts are able to dynamically guide the noise-conditioned denoising process. Therefore, we are able to efficiently train a unified denoising model for various count levels, and deploy it to different cases with personalized prompts. We evaluated on 1940 low-count PET 3D volumes with uniformly randomly selected 13-22\% fractions of events from 97 $^{18}$F-MK6240 tau PET studies. It shows our dual prompting can largely improve the performance with informed count-level and outperform the count-conditional model.

Dual Prompting for Diverse Count-level PET Denoising

TL;DR

The paper tackles the challenge of denoising low-count PET images across a continuous range of count levels. It introduces a dual-prompt framework combining a Specific Count Level Prior Prompt and a Learnable General PET Denoising Prompt , fused via cross-attention and injected into a 3D SwinUNet denoiser through skip connections. Key contributions include the explicit CLP mapping, the learnable GPD prompt, cross-attention fusion, and prompt injection strategy, all trained with an loss. Empirical results on F-MK6240 tau PET data show that the dual-prompt approach outperforms count-level conditioning and prior prompt methods, demonstrating improved MAE, MSE, PSNR, and 3D SSIM across 13–22% count levels and suggesting broader applicability to degradation-aware medical imaging tasks.

Abstract

The to-be-denoised positron emission tomography (PET) volumes are inherent with diverse count levels, which imposes challenges for a unified model to tackle varied cases. In this work, we resort to the recently flourished prompt learning to achieve generalizable PET denoising with different count levels. Specifically, we propose dual prompts to guide the PET denoising in a divide-and-conquer manner, i.e., an explicitly count-level prompt to provide the specific prior information and an implicitly general denoising prompt to encode the essential PET denoising knowledge. Then, a novel prompt fusion module is developed to unify the heterogeneous prompts, followed by a prompt-feature interaction module to inject prompts into the features. The prompts are able to dynamically guide the noise-conditioned denoising process. Therefore, we are able to efficiently train a unified denoising model for various count levels, and deploy it to different cases with personalized prompts. We evaluated on 1940 low-count PET 3D volumes with uniformly randomly selected 13-22\% fractions of events from 97 F-MK6240 tau PET studies. It shows our dual prompting can largely improve the performance with informed count-level and outperform the count-conditional model.
Paper Structure (10 sections, 4 equations, 3 figures, 1 table)

This paper contains 10 sections, 4 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Specific Count Level Prior (CLP) Prompt
  4. Learnable General PET Denoising (GPD) Prompt
  5. Dual Prompts Integration
  6. Prompt Injection and Training Protocol
  7. Experiments
  8. Conclusion
  9. COMPLIANCE WITH ETHICAL STANDARDS
  10. ACKNOWLEDGMENTS

Figures (3)

  • Figure 1: Illustration of applying (a) independent model xie2023unified, (b) conditional model he2021interactive, (c) blind prompting potlapalli2024promptir, and (d) our dual prompting for cross count-level PET denoising.
  • Figure 2: Illustration of our dual-prompt modules add-on to the 3D SWinUneter backbone.
  • Figure 3: Comparison of the denoising with different methods for a PET slice in (a) 20% count level and (b) 15% count level, respectively. The voxel-wise mean bias, standard deviation, and RMSE are the lower the better.