@article {Lee3191, author = {Kyungnam Lee and Arman Rahmim and Carlos Uribe}, title = {A Matlab-based Kinetic Modeling tool for fast and robust estimation of Patlak-based parameters with uncertainty information }, volume = {63}, number = {supplement 2}, pages = {3191--3191}, year = {2022}, publisher = {Society of Nuclear Medicine}, abstract = {3191 Introduction: Kinetic Modeling is a useful tool to extract biological or physiological information from the human body. Through compartment modeling, we can better understand not only the biological process of metabolism but also receptor-ligand bindings. Also, by analyzing the kinetic parameters characteristics, the results can be applied to diagnosis and therapy monitoring. Although there are dozens of established commercial or free software for kinetic modeling, there are still several points that need to be improved further to routinely perform kinetic modeling: 1) the calculation speed, 2) reporting of uncertainties for each estimated parameter, and 3) robust estimation.In this study, we aim to develop a kinetic modeling tool that tackles these three points using RANSAC-based GPU parallel processing algorithm.Methods: The kinetic modeling tool was developed using Matlab 2021b. Currently, the tool offers the Patlak model fitting with functionalities for the generation of parametric maps (i.e. slope and intercept) and provides uncertainties for each calculated parameter (i.e., relative errors). To improve the calculation speed, we implemented GPU parallel processing using the GPUfit open source library. The standard error is calculated using the Jacobian-based method. This was implemented directly by us as the GPUfit library did not include it. This allows the user to see and evaluate the estimated parameters in an objective and quantitative way; something missing in other kinetic modeling tools. Furthermore, to improve the robustness of our calculations, we implemented a RANSAC algorithm to remove outliers and improve the fitting process. Our toolset has been evaluated with a dataset of 6 whole body [18F]DCFPyL PET/CT scans which include a 6 minute acquisition of the chest (to measure the input function) and 16 subsequent whole-body passes that follow the pharmacokinetics for 90 minutes post-injection. The RANSAC algorithm{\textquoteright}s distance parameters was varied from d=2 to d=10 in intervals of 2. This was done to identify the effect of this parameter on the uncertainties of Ki and Vd. Results: For a volume comprised of 256 x 256 x 409 voxels, encompassing the whole-body, computational time was improved by a factor of 2200 compared to CPU-based framework. As a result, we can generate parametric maps with uncertainty information within 25 min for our [18F]DCFPyL cases. Times may vary depending on the platform and computer hardware available.In addition, by offering the functionalities for the calculation of uncertainties of estimated parameters, the user is presented with the uncertainty map for each parameter. This results in a total of 4 parametric maps being generated: Ki-map and Vd-map, and its corresponding uncertainties. By using the RANSAC algorithm, the user can minimize outlier{\textquoteright}s effects on the fits, and reduce relative errors within the whole-body volume. For example, when not utilizing the RANSAC algorithm, relative errors for Ki and Vd were 6.48\% and 79.8\%, respectively. These relative errors changed from 3.89\% to 3.39\% for Ki with d=10 and d=2, respectively. Similarly, the uncertainties in Vd changed from 22.8\% to 4.8\% with d=10 and d=2, respectively. Conclusions: Our Matlab-based kinetic modeling tool improves calculation speed and provides uncertainty in the Patlak-based kinetic parameters. By incorporating the RANSAC algorithm for outliers, the relative errors were reduced up to a factor of 4.8, with an optimal distance d value of 2. This is something that other kinetic modeling tools lack. We are currently working on developing the functionality for two tissue compartmental model.}, issn = {0161-5505}, URL = {https://jnm.snmjournals.org/content/63/supplement_2/3191}, eprint = {https://jnm.snmjournals.org/content}, journal = {Journal of Nuclear Medicine} }