DUAL PROMPTING FOR DIVERSE COUNT-LEVEL PET DENOISING

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Positron emission tomography (PET) imaging is challenged by diverse count levels in acquired volumes, impacting denoising model generalization.
• Existing denoising methods struggle to effectively handle the wide range of noise variations inherent in PET data.
• Unified models for PET denoising are needed to address varying count statistics without compromising image quality.

Purpose of the Study:

• To develop a generalizable PET denoising method using prompt learning that accommodates diverse count levels.
• To propose a novel dual-prompt strategy for guiding PET denoising in a unified model.
• To enhance the adaptability and performance of PET denoising models for varied clinical applications.

Main Methods:

• Employed prompt learning with explicitly count-level and implicitly general denoising prompts.
• Developed a prompt fusion module to integrate heterogeneous prompts.
• Implemented a prompt-feature interaction module to dynamically guide noise-conditioned denoising.
• Utilized a dataset of 1940 low-count 3D PET volumes from 97 ¹⁸F-MK6240 tau PET studies.

Main Results:

• The dual-prompting approach significantly improved PET denoising performance compared to count-conditional models.
• The unified model demonstrated effective generalization across different PET count levels.
• Personalized prompts allowed for tailored denoising based on specific PET study characteristics.

Conclusions:

• Dual prompting offers an effective strategy for generalizable PET denoising across varied count levels.
• The proposed prompt learning framework enables efficient training and deployment of unified PET denoising models.
• This method advances the application of AI in medical imaging by improving PET image quality and consistency.