Innovations in clinical PET image reconstruction: advances in Bayesian penalized likelihood algorithm and deep...

Area of Science:

• Medical Imaging
• Radiology
• Nuclear Medicine

Background:

• Current PET image reconstruction aims for high image quality and quantitative accuracy.
• Bayesian penalized likelihood (BPL) algorithms offer noise suppression and edge preservation.
• Deep learning (DL) methods have emerged as post-processing techniques for noise reduction.

Purpose of the Study:

• To review the technical principles of BPL and DL-based PET reconstruction.
• To summarize the clinical performance of these advanced reconstruction algorithms.
• To discuss image quality and quantitative accuracy considerations for PET imaging.

Main Methods:

• Review of Bayesian penalized likelihood (BPL) algorithms (e.g., Q.Clear, HYPER Iterative).
• Review of deep learning (DL) methods (e.g., SubtlePET, AiCE, uAI® HYPER DLR, Precision DL).
• Discussion of hybrid approaches like uAI® HYPER DPR integrating DL into iterative reconstruction.

Main Results:

• BPL algorithms provide robust noise suppression and edge preservation via regularization.
• DL methods effectively reduce noise and preserve lesion contrast, allowing reduced scan times or doses.
• Hybrid approaches combine iterative reconstruction with deep learning for enhanced performance.

Conclusions:

• Advanced PET reconstruction techniques, including BPL and DL, significantly improve image quality and quantitative accuracy.
• These methods support reduced radiation exposure or faster imaging protocols without compromising diagnostic confidence.
• Further understanding and implementation of these techniques are crucial for diverse PET imaging applications.