Population-based deep image prior for dynamic PET denoising: A data-driven approach to improve parametric...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Dynamic Positron Emission Tomography (PET) imaging is crucial for quantitative analysis, but noise significantly degrades image quality and parametric quantification.
• High noise levels in dynamic PET scans can obscure important diagnostic information and reduce the accuracy of kinetic parameter (K_i) estimations.

Purpose of the Study:

• To enhance the quality and quantitative accuracy of K_i images derived from noisy dynamic PET data.
• To introduce a novel deep learning-based denoising technique, Population-based Deep Image Prior (PDIP), for dynamic PET imaging.

Main Methods:

• Proposed PDIP, integrating population-based prior information into the Deep Image Prior (DIP) optimization framework.
• Generated population-based priors using a supervised denoising model trained on a large static PET dataset (100 studies) with a 3D U-Net architecture.
• Evaluated PDIP against Prompts-matched Supervised (PS) and conditional DIP (CDIP) models using dynamic PET data from 23 patients (25% and 100% counts).

Main Results:

• Both PS and CDIP models effectively reduced noise but caused over-smoothing and lesion removal.
• CDIP, using a single static image prior, introduced artifacts and inaccurate K_i values, particularly in the descending aorta.
• PDIP achieved comparable noise reduction to PS and CDIP while preserving small lesions and improving K_i predictions, especially for lesions.

Conclusions:

• PDIP offers a superior approach for denoising dynamic PET images, outperforming existing supervised and conditional DIP methods.
• The novel integration of population-based priors in PDIP enhances quantitative accuracy and lesion detectability in dynamic PET scans.
• PDIP demonstrates significant potential for improving the diagnostic utility of quantitative dynamic PET imaging.