# Deep learning-enhanced digital-BGO versus TOF PET/CT: comparative assessment of detection, quantitation, and overall image quality

## TL;DR

A new deep learning algorithm improves digital-BGO PET imaging to match TOF systems in detecting and quantifying lesions, while reducing scan time and maintaining image quality.

## Contribution

Demonstrates that deep learning can emulate TOF performance in PET imaging, enabling shorter scan times without sacrificing diagnostic accuracy.

## Key findings

- OMNI6R with PDL-High achieved non-inferior lesion detection sensitivity compared to TOF-equipped DMI5R.
- PDL-High improved quantification and image quality metrics across all lesion sizes.
- Clinical workflow and patient comfort were enhanced due to reduced scan time and improved image quality.

## Abstract

We evaluate the Omni Legend 32 cm (OMNI6R), a digital-BGO PET/CT using the deep learning (DL) based algorithm, Precision Deep Learning (PDL), emulating time-of-flight (TOF) enhancement and compare its performance to the TOF-equipped Discovery-MI 25 cm (DMI5R) in terms of detection sensitivity, quantification, and overall image quality.

Thirty patients were administered with an average single dose of 2 MBq/kg [18F]-FDG and were scanned consecutively on DMI5R first and on OMNI6R afterwards. Total scan duration on DMI5R and OMNI6R were 10 and 6 min, respectively. OMNI6R data were reconstructed using Bayesian Penalized Likelihood (BPL) algorithm with a beta of 650 and PDL-High setting. A total of 150 inserted synthetic lesions (ISL), ranging in size from 6 to 10 mm and exhibiting contrast levels between 3 and 15 relative to their initial background activity, were distributed across the cohort. Three readers blindly assessed detection sensitivity and quantification of these lesions. We tested a non-inferiority hypothesis based on the ISL true positive rate (TPR) and compared calculated recovery coefficients (RC) using SUVmean and SUVmax metrics of the detected ISL. Additionally, image quality, sharpness, conspicuity, noise characteristics, and diagnostic confidence were assessed as clinical quality indicators with a 5-point Likert scale on clinical images without ISL, using same beta as DMI5R and different PDL settings (None, High, Medium, Low).

TPR were 84.67% (95% CI 80.04–89.29%) and 84.44% (95% CI 77.76–91.13%) respectively for DMI5R and OMNI6R-PDL-High, and demonstrated non-inferiority. OMNI6R-PDL-High yielded higher RC without overestimation for all ISL sizes. Remarkably, these findings were observed despite a 9% activity decay in ISL and a 40% reduction in whole-body acquisition time. All PDL settings led to increased average median scores across clinical quality metrics, surpassing the DMI5R in most cases.

OMNI6R using PDL-High demonstrated non-inferior diagnostic performance compared to DMI5R, as evidenced by ISL detection sensitivity and quantitation. Importantly, the use of OMNI-PDL-High did not increase the risk of false-negative findings, despite reductions in activity and acquisition time. OMNI6R using PDL enhances overall image quality while improving clinical workflow and patient comfort. These results support DL-based enhancement algorithms as effective solutions for non-TOF PET imaging.

Trial registration number and date of registration: NCT05154877, December 13th 2021.