TY - JOUR T1 - <strong>Automatic segmentation with triple-factor non-negative matrix factorization method improves precision of myocardial blood flow quantification from dynamic <sup>82</sup>Rb PET</strong> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1710 LP - 1710 VL - 59 IS - supplement 1 AU - Hui Liu AU - Jing Wu AU - Jing-Yi Sun AU - Tung-Hsin Wu AU - Ramesh Fazzone-Chettiar AU - Stephanie Thorn AU - Albert Sinusas AU - Yi-Hwa Liu Y1 - 2018/05/01 UR - http://jnm.snmjournals.org/content/59/supplement_1/1710.abstract N2 - 1710Objectives: : Region of interest (ROI) extraction for time activity curve (TAC) calculation in dynamic PET usually has non-negligible inter-observer variability by manual image segmentation or is sensitive to the threshold setting in 3D ellipsoid fitting Methods: This study aimed to develop a triple-factor non-negative matrix factorization (TNMF) method using dynamic PET data for automatic segmentation of left ventricular (LV) cavity and myocardium to reduce manual labor and improve precision of myocardial blood flow (MBF) quantification in dynamic 82Rb PET. Methods: The principle of the TNMF method is that the dynamic data matrix can be factorized into a product of the template matrix of organs, the TAC matrix of organs and the system point spread function (PSF) matrix. To reduce computational cost, an initial coarse mask including LV cavity and myocardium was drawn manually on image. The three matrices could be solved successively within the same iteration, with the orthogonality constraint on the template matrix, circumventing potential template overlap among different organs. Normal NCAT phantom with LV myocardium, LV cavity, right ventricular (RV) myocardium and RV cavity was used for computer simulation. The TACs of LV and RV cavities were simulated with bi-exponential functions. The LV myocardial TAC was calculated using 1T model (K1 = 1 mL/min/cm3, k2 = 0.2/min). The values of TACs were assigned on a voxel-by-voxel basis in the NCAT phantom to generate a sequence of 27 dynamic PET images. Three count levels (100%, 10% and 1% of counts in a typical clinical 82Rb PET study) each with 30 realizations were simulated with Gaussian noise embedded in the dynamic images. A 3D Gaussian function with 7.5 mm FWHM mimicking the system PSF was convoluted with the dynamic images. After applying TNMF, the ROIs of LV cavity and myocardium were automatically extracted from the template matrix and used for TACs calculation. Then, K1 was derived from 1T fitting. For comparison, the fitting method with LV myocardium ROI from 3D ellipsoid fitting and a 3x3x3 cubic LV cavity ROI was adopted for calculations of TACs and K1. Correlation coefficient (CC) between TNMF extracted and true ROIs was calculated. Also, the CC of TAC between TNMF/fitting method and the truth was compared. K1 results were obtained with both methods. Additionally, TNMF was evaluated and compared with the fitting method using 18 healthy human subjects undergone both rest and stress dynamic 82Rb PET imaging. Results: In computer simulations with different count levels, TNMF provided robust ROI segmentation for LV myocardium and cavity with consistent PSF FWHM estimation, and the CCs of TAC for both ROIs were almost equal to 1. However, the CC of TAC was reduced with increased noise using the fitting method, presumably due to the complex threshold tuning. The same trend was also found in K1. Both methods with 100% counts resulted in excellent estimation of K1. The bias and standard deviation of K1 using the fitting method increased with noise, whereas TNMF showed stable performance even at the extreme low-count level (1%). For the rest human studies, the correlation of K1 (R2 = 0.73) showed good agreement between both methods, demonstrating the feasibility of the TNMF method used in patients. For the stress data, K1 correlation between these two methods was moderate (R2 = 0.51), which may be attributable to the severe partial volume effect (usually much smaller LV cavity and higher myocardium uptake in stress than in rest). The PSF FWHM obtained with TNMF (6.5±0.4 mm) was consistent with the system PSF reported in literatures. Conclusion: The proposed TNMF method is feasible for automatic segmentation of LV myocardium and cavity from dynamic 82Rb PET data in both simulation and clinical studies, with lower bias and standard derivation in MBF quantification compared to the conventional fitting method, especially in low-count/high-noise studies. View this table:The simulation results of the TNMF and fitting methods ER -