Abstract
Abstract
Objective
High sensitivity of long axial field-of-view (LAFOV) PET/CT systems enables diagnostic PET imaging with substantially reduced injected activity, increasing the relative contribution of CT to overall radiation exposure. Although multiple CT-free PET frameworks have been proposed, differences in evaluation methodology, including the use of scan-specific CT-information for precise patient-bed placement or for image registrations to handle PET-CT misalignments, complicate direct performance comparisons. This study performed a unified, CT-independent evaluation of CT-free frameworks for quantitative
18
F-FDG LAFOV PET imaging.
Methods
Six CT-free attenuation correction frameworks were investigated using the same non-curated patient dataset. Two attenuation-map (
μ
-map) generation approaches were first evaluated: a physics-based one in which transmission scans were performed using the intrinsic background radiation of lutetium oxyorthosilicate scintillators (LSO-Tx), eventually generating \({\mu}_{\mathrm{LSO\text{-}Tx}}\), and a data-driven approach in which a convolutional neural network (CNN) synthesized \({\mu}_{\mathrm{CNN}}\) from non-attenuation-corrected (NAC) PET image. Additional physics-based enhancements of both approaches were investigated by independently using the maximum likelihood-based reconstruction algorithms of activity and attenuation (MLAA) or attenuation correction factors (MLACF), generating the physics-based \({\mu}_{\mathrm{LSO\text{-}Tx}}^{\mathrm{MLAA}}\)and \({\mu}_{\mathrm{LSO\text{-}Tx}}^{\mathrm{MLACF}}\), and the data-physics hybrids of \({\mu}_{\mathrm{CNN}}^{\mathrm{MLAA}}\) and \({\mu}_{\mathrm{CNN}}^{\mathrm{MLACF}}\). Framework performances were evaluated using SUV
mean
extracted from attenuation-corrected PET images.
Results
CNN-based frameworks demonstrated substantially lower SUV
mean
biases than LSO-Tx-based approaches. Biases (mean ± standard deviation) for standalone, MLAA-, and MLACF-enhanced CNN frameworks were -9±7%, -12±13%, and -1.6±10.2%, respectively, compared with 29±29%, 26±100%, and 15±15% for the corresponding LSO-Tx frameworks. Low-attenuating objects located outside the patient body, including immobilization equipment and clothing, introduced median SUV
mean
biases of approximately 6%. Localized attenuation-map inaccuracies generated by CNN or LSO-Tx, such as missing internal air-pockets or metal implants, were partially corrected by MLAA and MLACF.
Conclusion
Among the investigated CT-free frameworks, the hybrid framework that combined NAC-PET-based deep learning with MLACF enhancement achieved the highest quantitative accuracy. The unified CT-independent evaluation further demonstrated that low-attenuating objects substantially influence PET quantification and should be explicitly considered in future development and evaluation of CT-free PET frameworks.