Scanner-integrated reconstruction versus post-processing deep learning for low-count ¹⁸F-FDG PET/CT: a comparative...

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Low-count positron emission tomography/computed tomography (PET/CT) imaging is crucial for many clinical applications.
• Acquisition time reduction in PET/CT is desirable to minimize patient discomfort and motion artifacts.
• Deep learning (DL) offers potential solutions for enhancing low-count PET/CT images.

Purpose of the Study:

• To compare two DL approaches for low-count PET/CT: deep progressive reconstruction (DPR) and deep-learning image-domain post-processing enhancement (POST).
• To evaluate the performance of DPR and POST in terms of image quality, quantitative metrics, and agreement with clinical reference standards.

Main Methods:

• Sixty-seven patients undergoing 18F-FDG PET/CT were included.
• Images were reconstructed using OSEM (clinical reference), DPR, and POST (RaDynPET).
• Image quality was assessed by nuclear medicine physicians, and quantitative metrics (SNR, TBR, CNR) were calculated. Non-inferiority versus OSEM was prespecified.

Main Results:

• Both DPR and POST achieved higher reader scores than time-matched OSEM.
• POST was superior at 30s, while DPR was superior at 60s acquisition times.
• DPR demonstrated monotonic increases in CNR and SNR, with TBR improvements, while POST gains attenuated at 60s.

Conclusions:

• Both DPR and POST significantly improved or preserved image quality, enabling scan-time reduction.
• POST is supported for 1/4 acquisition time, and DPR is favored from 1/2 time onward.
• Both methods demonstrated excellent agreement with the clinical reference OSEM standard.