Using long-axial field-of-view dynamic PET/CT data to generate synthetic contrast enhanced CT images

Abstract

ABSTRACT Purpose Dynamic whole-body [ 18 F]FDG PET/CT (DWB-PET/CT) imaging performed on long-axial field-of-view (LAFOV) systems enables the simultaneous assessment of tracer delivery and metabolism across extended anatomical regions. We investigated whether tracer delivery information derived from DWB-PET/CT scans could be used to generate biologically informed synthetic contrast-enhanced CT (ceCT) images using deep learning. Methods Ten patients underwent [ 18 F]FDG DWB-PET/CT and a co-registered ceCT. Voxelwise kinetic modelling was performed to derive unidirectional [ 18 F]FDG tissue clearance (K 1 ; ml g -1 min -1 ) maps which we hypothesise will correlate with CT contrast-media delivery. A deep-learning model was trained using K 1 maps and non-contrast enhanced CT (nceCT) images (for attenuation correction) as input to generate synthetic ceCT images (ceCT DWBPET-CT ). To evaluate the utility of incorporating DWB-PET/CT data, a deep-learning model was trained using only nceCT images as input (ceCT CTONLY ). Quantitative and task-orientated analyses were performed by comparing the simulated contrast enhancement for reference anatomical sites. Additionally, semi-quantitative consensus assessments were made for critical anatomical structures, image quality, and clinical usefulness of the images. Results The K 1 maps were used to obtain synthetic ceCT images which bore good visual fidelity to the ground-truth clinical ceCT images. The DWB-PET/CT data added greater value compared to a CT-only approach; the ceCT DWBPET-CT showed lower bias and lower mean percentage error compared to the ceCT CTONLY (median percentage error 2.79% vs. 8.41%) and preserved more anatomical structures with higher subjective image quality. The reference standard ceCT added additional clinical value compared to the nceCT images in 93% of cases, with the ceCT DWBPET-CT adding greater value compared to the ceCT CTONLY images (additional value added in 70% vs. 17% of cases). Conclusion We demonstrate the feasibility of using DWB-PET/CT from LAFOV to train a DL-model for generating synthetic ceCT images. The method is a potential means of simulating ceCT in cases where iv contrast is contraindicated or unavailable.