Enhanced PET imaging using progressive conditional deep image prior.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Unsupervised learning improves positron emission tomography (PET) image quality with limited data.
• Direct unsupervised learning struggles with large input-to-target image gaps, potentially reducing lesion detectability.

Purpose of the Study:

• To develop a novel unsupervised learning method for enhanced lesion detectability in PET patient studies.
• To address challenges in unsupervised learning when significant differences exist between input and target PET images.

Main Methods:

• A deep progressive learning strategy was employed, decomposing one-step unsupervised learning into two stages.
• The first network uses anatomical images; the second uses low-noise PET images and the first network's output as prior information for iterative reconstruction.

Main Results:

• The proposed progressive unsupervised method demonstrated superior performance compared to non-deep learning and standard unsupervised methods.
• Its effectiveness was validated in both phantom and patient studies, achieving results comparable to supervised learning methods.

Conclusions:

• A progressive unsupervised learning approach effectively improves PET image noise performance and lesion detectability.
• This method offers a viable alternative to supervised learning when large datasets are unavailable.