@article {Wang1733, author = {Peipei Wang and Yuru He and Arman Rahmim and Wufan Chen and Lijun Lu}, title = {Generalized factor analysis incorporating alpha-divergence and kinetics-based clustering: application to dynamic myocardial perfusion PET imaging}, volume = {59}, number = {supplement 1}, pages = {1733--1733}, year = {2018}, publisher = {Society of Nuclear Medicine}, abstract = {1733Objectives: We aimed to improve quantitative accuracy of time-activity curves (TACs) of regional tissue from dynamic myocardial PET images. To this end, we proposed and assessed a generalized factor analysis method incorporating α-divergence measure and kinetics-based clustering. Methods: Conventional factor analysis (such as the minimal structure overlap (MSO) method) commonly models the noise distribution of dynamic PET images as independently Gaussian or Poisson to define the objective function that is optimized (e.g. leading to least squares model of divergence between measured and fit data for the Gaussian model). Additionally, uniqueness constraints are incorporated as prior to minimize structure overlap (i.e. between blood and myocardium signals). However, the noise distribution in the reconstructed image domain is not accurately described by independent distributions. Therefore, we propose use of a generalized factor analysis (denoted as GFA-KB) wherein α-divergence was incorporated as a measure to assess the difference between the estimated model and measured data. Furthermore, kinetic-based clustering was incorporated into the uniqueness constraint as a prior. Using simulated myocardial perfusion 82Rb PET data, we first decomposed the dynamic images as initial factor curves and factor images by minimizing the α-divergence, then used the fuzzy c-means framework on dynamic PET images to derive the kinetics-based cluster, finally incorporating the prior information into non-uniqueness to derive the factor curves and factor images. Results: The experimental results show that the generalized factor analysis performs better than the traditional model in accuracy and sensitivity. Specifically, with the noise level increasing, root mean square error (RMSE) of blood factor with MSO and GFA-KB methods increased from 2.9\% to 9.0\% versus 0.64\% to 1.3\% respectively. Meanwhile, RMSE of myocardium factor with MSO increased from 4.4\% to 11.3\%, while with GFA-KB, it decreased from 1.4\% to 0.47\%. Furthermore, residue percent of α-divergence when varying the α value was lower than that for least squares fitting. Conclusions: The proposed generalized factor analysis enhances quantitative accuracy of tissue TACs compared with conventional factor analysis method by incorporating kinetics-based clustering and optimal α-divergence. Acknowledgments: This work was supported by the National Natural Science Foundation of China under grants 61628105, 81501541, the National key research and development program under grant 2016YFC0104003, the Natural Science Foundation of Guangdong Province under grants 2016A030313577, and the Program of Pearl River Young Talents of Science and Technology in Guangzhou under grant 201610010011.}, issn = {0161-5505}, URL = {https://jnm.snmjournals.org/content/59/supplement_1/1733}, eprint = {https://jnm.snmjournals.org/content}, journal = {Journal of Nuclear Medicine} }