Linear ridge regression with spatial constraint for generation of parametric images in dynamic positron emission tomography studies

Citation
Y. Zhou et al., Linear ridge regression with spatial constraint for generation of parametric images in dynamic positron emission tomography studies, IEEE NUCL S, 48(1), 2001, pp. 125-130
Citations number
15
Categorie Soggetti
Apllied Physucs/Condensed Matter/Materiales Science","Nuclear Emgineering
Journal title
IEEE TRANSACTIONS ON NUCLEAR SCIENCE
ISSN journal
00189499 → ACNP
Volume
48
Issue
1
Year of publication
2001
Part
1
Pages
125 - 130
Database
ISI
SICI code
0018-9499(200102)48:1<125:LRRWSC>2.0.ZU;2-A
Abstract
Due to its simplicity, computational efficiency, and reliability, weighted linear regression (WLR) is widely used for generation of parametric imaging in positron emission tomography (PET) studies, but parametric images estim ated by WLR usually have high image noise level. To improve the stability a nd signal-to-noise ratio of the estimated parametric images, we have added ridge regression, a statistical technique that reduces estimation variabili ty at the expense of a small bias. To minimize the bias, spatially smoothed images obtained with WLR are used as a; constraint for ridge regression. T his new algorithm consists of two steps. First, parametric images are gener ated by WLR and are spatially smoothed. Ridge regression is then applied us ing the smoothed parametric images obtained in the first step as the constr aint. Since both "generalized" ridge regression. and "simple" ridge regress ion are used in statistical applications, we evaluated specifically in this study the relative advantages of the two when incorporated for generating parametric images fi om dynamic O-15 water PET studies. Computer simulation s of a dynamic PET study with the spatial configuration of Hoffman's brain phantom and a real human PET study were used as the data for the evaluation . Results reveal ridge regressions improve image quality of parametric imag es for studies with high or middle noise level. as compared to WLR. Use of generalized ridge regression offers little advantage over that of simple ri dge regression.