UC Davis

Dynamic positron emission tomography (PET) imaging captures a series of PET images over time and monitors the spatiotemporal distribution of the radiotracer administered to the body. Tracer kinetic modeling and parametric imaging (i.e., voxel-wise kinetic modeling) are a technique for dynamic PET imaging. It enables the quantification of kinetic parameters via the mathematical modeling of the time-varying tracer distribution. The quantified parameters represent the tracer kinetics and can potentially serve as biomarkers for various diseases. However, the development and application of kinetic modeling are largely limited by the short axial field-of-view (AFOV) (15-30 cm) of conventional PET scanners. This short AFOV not only restricts the anatomical coverage of the body but also confines the temporal resolution of dynamic scans to typically 10-40s/frame due to the low detection sensitivity. The introduction of total-body PET systems, such as the 194-cm long uEXPLORER, enables the total-body field of view and significantly increases the detection sensitivity. Propelled by these advancements, we developed kinetic modeling with total-body PET in multiple aspects, emphasizing applications to lung disease but also broadly encompassing systemic disease. First, we investigated the high temporal resolution (HTR) kinetic modeling by leveraging HTR dynamic imaging (e.g., 1s/frame) with the total-body PET scanner. Second, multi-organ kinetic modeling was studied, taking advantage of the simultaneous imaging of the entire body. Third, deep learning was explored to pursue efficient approaches for total-body parametric imaging. The investigation of HTR kinetic modeling in this study focuses on the lung, an organ unique for its dual blood supplies from the right ventricle and the left ventricle. The HTR dynamic imaging enables the capture of the rapid-changing early kinetics of the lung and its two blood supplies. However, existing kinetic models are insufficient for modeling the acquired HTR data. Hence, we first studied necessary corrections to the right ventricle input function, which is the dominant blood supply to the normal lung tissue. Corrections of time delay and dispersion were demonstrated to largely improve model fitting and impact the lung kinetic parameter quantification, leading to more reasonable estimates of fractional blood volume vb (~0.14) and the detected aging effect of vb, both within physiological expectations. Second, considering that lung tumors can have altered blood supplies compared with normal lung tissue, we proposed the dual-blood input function (DBIF) for lung kinetic modeling. The DBIF further improved the fitting quality, especially for lung tumors. In addition, the left ventricle supply fraction f that is uniquely quantified by the DBIF model was significantly higher in lung tumors (~0.3) than in normal lung tissue (~0.04). Besides the HTR, total-body dynamic PET also permits the kinetic quantification of multiple organs and multiple parameters, which is promising for the evaluation of systemic diseases. In this work, we applied multi-organ kinetic modeling to evaluate metabolic changes in coronavirus disease 2019 (COVID-19) recovery. A higher lung 18F-fluorodeoxyglucose (FDG) net influx rate Ki and a higher bone marrow FDG delivery K1 were detected in recovering COVID-19 subjects compared to healthy subjects with statistical significance. These multiparametric findings may be associated with continued inflammation during the COVID-19 recovery and will be otherwise missed if only assessed with the standardized uptake value (SUV) using whole-body static PET imaging. While conventional kinetic modeling methods can be time-consuming for total-body parametric imaging owing to the large data amount to process, deep learning is promising for providing more efficient approaches. Hence, our study investigated the application of deep learning for total-body parametric imaging. The first study focused on total-body kinetic model selection, which aims to identify the appropriate kinetic model for body voxels and suppress artifacts in parametric images. We proposed a single-subject deep learning strategy to avoid the need for a population database for model training, and our preliminary tests showed the proposed method achieved better efficiency than the commonly used model selection method. In the second study, we explored deep learning for total-body voxel-wise parameter quantification. We proposed the Deep Patlak, a deep neural network method for the estimation of net influx rate Ki with its architectural design inspired by the conventional Patlak plot. The proposed Deep Patlak decreased the time cost for total-body parametric imaging of Ki than the conventional model-fitting-based method, while it is also more interpretable as compared to alternative neural network models. The parametric image by Deep Patlak showed good potential in imaging lung metastases. In summary, this dissertation investigated tracer kinetic modeling and parametric imaging with total-body PET and its applications to lung disease and beyond from different angles, including high temporal resolution kinetic modeling, multi-organ kinetic modeling, and deep learning for total-body parametric imaging. We demonstrated the feasibility of high temporal resolution kinetic modeling and the potential for disease evaluation utilizing the rapid-changing early kinetics. The multi-organ kinetic modeling enables a multiparametric quantification and assessment of the tracer kinetics in the entire body. The deep learning studies contribute to enhancing the effectiveness and efficiency of total-body parametric imaging. Our investigations highlight the combination of tracer kinetic modeling and total-body dynamic PET imaging in various contexts, demonstrating it as a sensitive tool to evaluate the human body, in both health and disease.