A Fast Nonlinear Method for Parametric Imaging of Myocardial Perfusion by Dynamic 13N-Ammonia PET

Journal of Nuclear Medicine

Heinrich R. Schelbert¹  S. Raymond Golish¹  Sanjiv S. Gambhir¹  Jens D. Hove¹ 

[1] Crump Institute for Molecular Imaging and Departments of Computer Science, Molecular and Medical Pharmacology, and Biomathematics, University of California, Los Angeles, Los Angeles, California; and Rigshospitalet, Copenhagen, Denmark Crump Institute for Molecular Imaging and Departments of Computer Science, Molecular and Medical Pharmacology, and Biomathematics, University of California, Los Angeles, Los Angeles, California; and Rigshospitalet, Copenhagen, Denmark Crump Institute for Molecular Imaging and Departments of Computer Science, Molecular and Medical Pharmacology, and Biomathematics, University of California, Los Angeles, Los Angeles, California; and Rigshospitalet, Copenhagen, Denmark

关键词: parametric imaging;    ¹³N-ammonia PET;    sigmoidal networks;    artificial neural networks;    function estimation;    nonlinear regression;   

DOI  :  

学科分类：医学（综合）

来源: Society of Nuclear Medicine

【 摘 要 】

A parametric image of myocardial perfusion (mL/min/g) is a quantitative image generated by fitting a tracer kinetic model to dynamic 13N-ammonia PET data on a pixel-by-pixel basis. There are several methods for such parameter estimation problems, including weighted nonlinear regression (WNLR) and a fast linearizing method known as Patlak analysis. Previous work showed that sigmoidal networks can be used for parameter estimation of mono- and biexponential models. The method used in this study is a hybrid of WNLR and sigmoidal networks called nonlinear regression estimation (NRE). The purpose of the study is to compare NRE with WNLR and Patlak analysis for parametric imaging of perfusion in the canine heart by 13N-ammonia PET. Methods: A simulation study measured the statistical performance of NRE, WNLR, and Patlak analysis for a probabilistic model of time–activity curves. Four canine subjects were injected with 740 MBq 13N-ammonia and scanned dynamically. Images were reconstructed with filtered backprojection and resliced into short-axis cuts. Parametric images of a single midventricular plane per subject were generated by NRE, WNLR, and Patlak analysis. Small regions of interest (ROIs) were drawn on each parametric image (8 ROIs per subject for a total of 32). Results: For the simulation study, the median absolute value of the relative error for a perfusion value of 1.0 mL/min/g was 16.6% for NRE, 17.9% for WNLR, 19.5% for Patlak analysis, and 14.5% for an optimal WNLR method (computable by simulation only). All methods are unbiased conditioned on a wide range of perfusion values. For the canine studies, the least squares line fits comparing NRE (y) and Patlak analysis (z) with WNLR (x) for all 32 ROIs were y = 1.02x − 0.028 and z = 0.90x + 0.019, respectively. Both NRE and Patlak analysis generate 128 × 128 parametric images in seconds. Conclusion: The statistical performance of NRE is competitive with WNLR and superior to Patlak analysis for parametric imaging of myocardial perfusion. NRE is a fast nonlinear alternative to Patlak analysis and other fast linearizing methods for parametric imaging. NRE should be applicable to many other tracers and tracer kinetic models.

【 授权许可】

Unknown