Single-Subject Deep-Learning Image Reconstruction With a Neural Optimization Transfer Algorithm for PET-Enabled...

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Dual-energy computed tomography (DECT) combined with positron emission tomography (PET) has clinical potential but faces hardware and radiation dose challenges.
• Existing PET-enabled DECT methods reconstruct gamma-ray CT (gCT) images from PET data but haven't fully utilized prior knowledge.
• Deep learning offers promise but requires large datasets, impractical for novel methods like PET-enabled DECT.

Purpose of the Study:

• To develop a single-subject deep learning method for improving gCT image reconstruction in PET-enabled DECT.
• To enhance gCT image quality and multi-material decomposition without population-based pre-training.

Main Methods:

• Proposed a single-subject method using neural-network representation as a deep coefficient prior for gCT image reconstruction.
• Formulated the problem as tomographic estimation of nonlinear neural-network parameters.
• Employed an optimization transfer strategy with quadratic surrogates, including PET activity update, gCT update, and least-square neural-network learning.

Main Results:

• The proposed neural optimization transfer algorithm demonstrated monotonic increase in data likelihood.
• Significant improvements in gCT image quality were observed compared to existing methods.
• Enhanced multi-material decomposition capabilities were validated using simulation, phantom, and patient data.

Conclusions:

• The single-subject deep learning approach effectively improves gCT image reconstruction for PET-enabled DECT.
• This method overcomes the limitations of large database requirements for deep learning in novel imaging techniques.
• The technique offers a promising solution for advancing DECT applications on PET/CT scanners.