TY - JOUR T1 - A machine learning-based parametric imaging algorithm for noninvasive quantification of dynamic [68Ga]DOTATATE PET-CT JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1186 LP - 1186 VL - 60 IS - supplement 1 AU - Yun Zhou AU - Jiangyuan Yu AU - Min Liu AU - Hebei Li AU - Zhi Yang AU - Richard Wahl Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/1186.abstract N2 - 1186Objectives: [68Ga]DOTATATE PET-CT are now routinely used clinically for imaging neuroendocrine tumor (NET) somatostatin receptor subtype 2 (SSTR2) density. Clinical studies show that [68Ga] DOTATATE PET-CT are complementary to FDG PET-CT for improving the accuracy in NET detection, characterization, grading, staging, and predicting/monitoring the responses to treatments. As peptide receptor radionuclide therapy in NET is now widely available, the importance of noninvasive quantitative [68Ga]DOTATATE PET-CT NET imaging is increasingly recognized. The objective of this study is to evaluate a machine learning-based parametric imaging algorithm for noninvasive quantification of [68Ga]DOTATATE PET-CT imaging to allow more sophisticated and less time-dependent uptake measures than SUV. Methods: Nineteen [68Ga]DOTATATE 45-min dynamic PET-CT NET patient scans were collected. Regions of interests (ROIs) were manually defined on the fused PET-CT images and applied to dynamic PET images to generate ROI time activity curves (TACs, CPET(t)). Aortic TACs were used to estimate the blood input function for kinetic modeling. Random forests regression (1) machine learning model (RFM) was proposed to estimate tracer uptake rate constants Ki and initial distribution volume (Vid) directly from tissue TACs without input function. The Ki and Vid obtained by graphical analysis using Patlak plot with blood input function were used as reference. The RFM algorithm was firstly evaluated by computer simulation. An irreversible 2-tissue compartmental model was used to fit the ROI TACs. Gaussian noise with zero mean and variance of αCPET(t)exp(0.693ti/λ)/Δti was added to the fitted ROI TACs for simulation of different noise levels from ROI to voxelwise kinetics, where α is from 0.05 to 0.4 for noise level, λ is tracer physical half-life, and Δti is the duration of dynamic PET scan at frame i. The RFMs trained by the fitted ROI TACs were then applied to voxelwise TACs to generate Ki and Vid parametric images. Results: The starting time point of linear regression for Patlak plot was as early as 10 min post tracer injection. The variances of estimates of Ki and Vid from RFM are almost constant over all noise levels. In contrast, the variances of estimates from the Patlak plot (t[asterisk] = 35 min) at highest noise level (α = 4) are 7 and 16 times higher than the ones at low noise level (α = 0.05) for Ki and Vid, respectively. Consistent with simulation results, the Ki and Vid images generated by the RFM from the last 10 min are comparable to the ones from last 35 min scan. The Ki and Vid images generated by the Patlak plot from the last 10 min are not reliable and appear to be too noisy for clinical application. The high image quality of Vid images generated by the RFM may provide additional quantitative measurements related to tumor blood volume and vasculature for clinical assessment. Conclusions: The Random Forests regression machine learning model is robust to generate parametric images of Ki and Vid. The clinical value of Ki and Vid parametric images generated from RFM method is under investigation. References 1. Breiman L. Random Forests. Machine Learning. 2001;45(1):5-32. doi: 10.1023/a:1010933404324. ER -