Projection Space Implementation of Deep Learning-Guided Low-Dose Brain PET Imaging Improves Performance over...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Nuclear Medicine

Background:

• Low-dose (LD) PET imaging is crucial for reducing radiation exposure but often compromises image quality.
• Deep learning techniques offer potential for reconstructing high-quality PET images from limited data.
• Synthesizing full-dose (FD) PET images from LD acquisitions is essential for maintaining diagnostic accuracy.

Purpose of the Study:

• To evaluate the performance of deep learning-based full-dose (FD) PET image synthesis from low-dose (LD) PET data.
• To compare synthesis in image space (PIS) versus sinogram space (PSS) for diagnostic quality preservation.
• To assess the utility of synthesized PET data for quantitative analysis, including radiomic features.

Main Methods:

• Retrospective analysis of 140 clinical brain PET/CT studies.
• Simulation of LD PET data by randomly selecting 5% of events from FD list-mode data.
• Implementation of a modified 3D U-Net model to predict FD sinograms (PSS) and images (PIS) from LD data.
• Qualitative assessment by nuclear medicine specialists and quantitative analysis using PSNR, SSIM, SUV bias, and radiomic features.

Main Results:

• Nuclear medicine specialists rated all PSS images as good to excellent (score ≥ 4).
• PIS achieved PSNR of 0.96 ± 0.03 and SSIM of 0.97 ± 0.02; PSS achieved higher values (31.70 ± 0.75 and 37.30 ± 0.71, respectively).
• PSS demonstrated superior performance with lower average SUV bias (0.24% ± 0.96%) and variance compared to PIS (1.05% ± 1.44%).
• Radiomic feature analysis showed lower relative error for PSS (-1.07 ± 1.77 for PIS vs. 0.28 ± 1.4 for PSS).

Conclusions:

• Deep learning-based synthesis of FD PET data from LD acquisitions is feasible without sacrificing diagnostic quality.
• Synthesis in sinogram space (PSS) yields superior image quality and quantitative accuracy compared to image space (PIS).
• The PSS method shows significant potential for improving LD PET image reconstruction and analysis in clinical practice.