Effects of a deep learning-based image quality enhancement method on a digital-BGO PET/CT system for F-FDG whole-body examination.

BACKGROUND

The digital-BGO PET/CT system, Omni Legend 32, incorporates modified block sequential regularized expectation maximization (BSREM) image reconstruction and a deep learning-based time-of-flight (TOF)-like image quality enhancement process called Precision DL (PDL). The present study aimed to define the fundamental characteristics of PDL using phantom and clinical images.

METHODS

A NEMA IEC body phantom was scanned using the Omni Legend 32 PET/CT system. All PET/CT images were acquired over 60 and 90 s per bed position, with a 384 × 384 matrix. Phantom images were reconstructed using OSEM + PSF and BSREM at β values of 100-1,000, combined with low (LPDL), medium (MPDL), and high (HPDL) PDL. We evaluated contrast recovery, background variability, and the contrast-to-noise ratio (CNR) of a 10 mm hot sphere. Thirty clinical whole-body F-FDG PET/CT examinations were included. Clinical images were reconstructed using OSEM + PSF and BSREM at β values of 200, 300, 400, 500, and 600, determined based on findings from the phantom study, combined with the three PDL models. Noise levels, mean SUV (SUV), and the signal-to-noise ratio (SNR) of the liver as well as signal-to-background ratios (SBR) and maximum SUV (SUV) of lesions were evaluated. Two blinded readers evaluated visual image quality and rated several aspects to complement the analysis.

RESULTS

Contrast recovery and background variability decreased as the β value increased. This trend was consistent even when PDL processing was added to BSREM. Increased strength of the PDL models led to higher CNR. Noise levels decreased as a function of increasing β values in BSREM, resulting in a higher SNR, but lower SBR. Combining PDL with BSREM resulted in all β values producing better results in terms of noise, SBR, and SNR than OSEM + PSF. As the PDL increased (LPDL < MPDL < HPDL), noise levels, SBR, and SNR became higher. The β values of 400, 200, 300, and 300 for BSREM, LPDL, MPDL, and HPDL, respectively, resulted in noise equivalent to OSEM + PSF but significantly increased the SUV (9%, 15%, 18%, and 27%), SBR (16%, 17%, 20%, and 32%), and SNR (17%, 19%, 31%, and 36%), respectively. The visual evaluation of image quality yielded similar scores across BSREM + PDL reconstructions, although BSREM with β = 600 combined with MPDL delivered the best overall image quality and total mean score.

CONCLUSION

The combination of BSREM and PDL significantly enhanced the SUV of lesions and image quality compared with OSEM + PSF. A combination of BSREM at β values of 500-600 and MPDL is recommended for oncological whole-body PET/CT imaging when using PDL on the Omni Legend.