CT-less deep learning-based parametric imaging with image-derived input function

Various approaches for CT-less PET acquisition have been proposed and evaluated in the standalone PET scanners or low-dose clinical applications. CT-less parametric imaging has yet been investigated and poses challenges in terms of both attenuation correction and estimation of image-derived input function (IDIF). In this work, we proposed a CT-less deep learning-based solution of parametric imaging with IDIF. A synthesized attenuation map from non-attenuation-corrected (NAC) PET was used for the dynamic PET image reconstruction, as well as the segmentation of aorta mask and estimation of IDIF. Clinical static
¹⁸
F-FDG PET/CT data were acquired for the training and evaluation of the synthesis network and the segmentation network, and dynamic whole-body
¹⁸
F-FDG scans were used to test the proposed method. High quantitative accuracy was found in the synthesized attenuation maps, segmented aorta masks, and estimated IDIFs. Small difference was observed in the subsequent Patlak parametric images as compared to the conventional CT-based results. The proposed CT-less solution was demonstrated feasible and accurate in the parametric imaging.