Diffusion-Based Model for Parametric <i> K _i </i> Generation From

Dynamic positron emission tomography (PET) parametric imaging typically requires a 60-minute acquisition period, causing patient discomfort and reducing clinical efficiency. This study explores the feasibility of generating parametric Ki images from 10-minute dynamic PET images acquired in the early or late scanning phases employing a multi-channel feature fusion cold sampling (MCFFCoS) framework. PET data from 103 patients are acquired using the uEXPLORER total-body PET/CT scanner during 60-minute scans. This study conducts deep learning experiments, taking early-phase or late-phase PET images as input, respectively. The generated Ki images are evaluated by visual quality and quantitative metrics, including root mean squared error (RMSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). Volumes of interest (VOIs) analysis is performed using linear regression and Bland-Altman plots. In the quantitative evaluation of total-body data, the parametric Ki images generated from late-phase PET data generally outperform those derived from early-phase data. The analysis of VOIs indicates that the appropriate scanning protocol for PET parametric imaging may vary for different body regions. The deep learning approach is able to generate high-quality parametric Ki images from 10-minute dynamic PET scans, bypassing the requirements of long acquisition time for the estimation of blood input function in kinetic modeling.