Quantitative comparison of CT-free frameworks of PET imaging in LAFOV PET/CT systems

Objective In long axial field-of-view (LAFOV) PET/CT systems, where a fraction of the standard injected activity can be adequate for PET imaging, (low-dose) CT can have the largest contribution to the radiation dose. Aiming to eradicate this radiation contribution, e.g., to reduce the risk for patients with long life expectancy, this study assesses the quantitative performance of CT-free frameworks for ¹⁸ F-FDG LAFOV PET/CT. Methods Six frameworks for generating linear attenuation coefficient maps ( µ -maps) were investigated. First, we used an AI method to synthesize µ -maps (\({\mu}_{\text{C}\text{N}\text{N}}\)) from non-attenuation corrected (NAC) PET images. Secondly, by exploiting the intrinsic background radiation of lutetium oxyorthosilicate scintillators, we performed transmission scans (LSO-Tx) to generate additional µ -maps (\({\mu}_{\text{L}\text{S}\text{O}\text{-}\text{T}\text{x}}\)). µ -map enhancement by joint activity and attenuation reconstruction algorithms was investigated independently by using either \({\mu}_{\text{C}\text{N}\text{N}}\) or \({\mu}_{\text{L}\text{S}\text{O}\text{-}\text{T}\text{x}}\) as initial conditions in maximum likelihood estimation algorithms of activity and attenuation (MLAA) or of activity and attenuation correction factors (MLACF). Performance assessment of each framework was based on µ -values and standardized uptake values (SUV) extracted from µ -maps and their corresponding attenuation-corrected PET images. Results With respect to CT-based µ -maps, the standalone CNN and LSO-Tx based frameworks yielded mean ± sd relative biases of 0.8 ± 9% and 18 ± 39%, respectively. By using MLAA with \({\mu}_{\text{C}\text{N}\text{N}}\) and \({\mu}_{\text{L}\text{S}\text{O}\text{-}\text{T}\text{x}}\) as priors, biases were measured at -10 ± 25% and 3 ± 28%, respectively, while their MLACF equivalents were measured at -1.7 ± 11.7% and 4 ± 11%. In SUV quantification, CNN-based frameworks yielded − 9 ± 7%, -13 ± 13%, and − 1.8 ± 10%, respectively for standalone, MLAA, and MLACF-enhanced frameworks, while the equivalent measures for the frameworks based on LSO-Tx yielded 28 ± 28%, 26 ± 98%, and 14 ± 15%. Conclusion From the six CT-free frameworks investigated, the combination of an AI-based approach using NAC-PET with MLACF-enhancement was found to be most promising, providing quantitative PET data with an average accuracy within 10% and/or within − 0.4 SUV.