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.
Introduction
Positron Emission Tomography (PET) is a vital diagnostic tool, but it often requires injecting patients with radioactive tracers. To minimize radiation exposure, clinicians prefer "low-count" (low-dose) scans. However, fewer photon counts lead to significantly higher noise, which can obscure critical metabolic details.
The technical challenge lies in the diversity of these scans. A denoising model trained for a 10% count level might fail on a 20% count level. While one could train a separate model for every possible dose, it is computationally expensive and impractical. Conversely, "all-in-one" models often struggle to generalize across different noise intensities, suffering from "degradation vanishing" where the model loses track of the specific noise level it is supposed to fix. This paper addresses this by using prompt learning to guide the denoising process dynamically.
Methodology
The authors introduce a Dual Prompting scheme integrated into a 3D Swin-UNet architecture. The "dual" nature comes from two distinct types of information:
- Specific Count Level Prior (CLP) Prompt: An explicit prompt that takes the known count level (a continuous scalar value $\delta$) and passes it through a Multi-Layer Perceptron (MLP) to inform the model of the expected noise intensity.
- Learnable General PET Denoising (GPD) Prompt: An implicit prompt consisting of learnable parameters that capture general knowledge about PET noise characteristics.
These prompts are merged using a Cross-Attention mechanism, ensuring the model pays attention to the right features at the right time. This fused prompt is then injected into the skip connections of the U-Net, acting as a set of instructions that guide the features as they move through the network.
Experimental Setup
The study utilized a robust dataset of 1940 low-count PET 3D volumes:
- Source: 97 $^{18}$F-MK6240 tau PET studies from 60 participants scanned on a GE DMI PET-CT system.
- Simulation: Full-count data was randomly subsampled to represent 13% to 22% of the original events.
- Reference: Full-count reconstructions served as the ground truth for training.
Results and Discussion
The Dual Prompting method outperformed all baselines across standard metrics. By treating the count level as a continuous variable, the model is significantly more flexible for real-world clinical settings where doses vary.
Key quantitative results include:
- PSNR: Achieved 19.99 dB, compared to 19.09 dB for a standard conditional model and 13.99 dB for a basic UNet without prompts.
- MAE: Reduced to 0.2205, significantly lower than the unprocessed low-count images (0.7132).
- Visual Quality: Qualitative results show that the dual prompt method produces cleaner images with fewer artifacts in the "Mean Bias" and "RMSE" maps compared to using only a single prompt or no prompt at all.
Conclusion
This work demonstrates that dual prompting is a powerful tool for generalizable PET denoising. By splitting the task into intensity-specific and noise-general components, the model achieves superior performance across varying count levels. Future research will focus on extending this framework to other tracers and imaging modalities like CT or MRI.
Summary updated