4D deep image prior: dynamic PET image denoising using an unsupervised four-dimensional branch convolutional neural...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Supervised convolutional neural networks (CNNs) excel at denoising positron emission tomography (PET) images but require extensive, high-quality datasets.
• Deep Image Prior (DIP) offers an unsupervised alternative, denoising images using only the target image itself.

Purpose of the Study:

• To develop an innovative unsupervised method for denoising dynamic PET images using a 4D CNN architecture within a DIP framework.
• To enable end-to-end dynamic PET image denoising without the need for large, paired datasets.

Main Methods:

• Proposed a novel 4D CNN architecture with a feature extractor and reconstruction branch for each time frame.
• Implemented an end-to-end training procedure for dynamic PET image denoising using the DIP approach.
• Utilized a subject's static PET image as auxiliary information, with dynamic PET images serving as training labels.

Main Results:

• The 4D DIP framework demonstrated superior quantitative and qualitative denoising performance compared to 3D DIP and other unsupervised methods.
• Evaluated using both simulated [18F]fluoro-2-deoxy-D-glucose (FDG) data and preclinical [18F]FDG and [11C]raclopride data.
• Successfully denoised dynamic PET images without requiring large, high-quality patient datasets.

Conclusions:

• The proposed 4D DIP framework offers a promising unsupervised solution for dynamic PET image denoising.
• This approach eliminates the need for extensive supervised training datasets, making it more accessible.
• The method shows significant potential for improving the quality of dynamic PET imaging analysis.