AbstractID: 6910 Title: The N-Dimensional Image Noise-Power Spectrum
Estimation of the image noise-power spectrum (NPS) is fundamental to objective characterization of the
performance of novel imaging systems. Advanced digital x-ray imaging technologies (e.g., flat-panel
imagers) are under development for numerous applications, including: projection radiography [2-D data
in the spatial domain (x,y)]; real-time fluoroscopy [2-D data in (x,y) as a function of time, t – essentially
constituting 3-D data]; and cone-beam CT [3-D data in (x,y,z)]. Accurate estimation of the NPS in each
case requires that the analysis knowledgeably account for signal correlation in each of the n dimensions.
For example, it is well known that analysis of a central slice of the 2-D NPS (e.g., using a “synthesized
slit”) must account for correlation (blur) in the orthogonal direction. Analogously, the 2-D NPS (e.g.,
from fluoroscopic data) is affected by correlation (image lag) in the temporal domain, and failure to
account for such results in underestimation of NPS (overestimation of DQE). Similarly, the 3-D NPS in
cone-beam CT depends on correlation (blur) in the 3-D spatial domain, which is known to be highly
asymmetric. This paper presents on a general framework for NPS analysis in n dimensions, highlighting
the important aspects of n-D correlation and sampling (i.e., n-D aliasing of the NPS). The issues of NPS
convergence and stationarity are addressed, and a general form for NPS normalization is given. The
framework is applied to real and simulated n-D image data, reducing to the familiar cases of 1-D and 2D NPS analysis, but extending generally to 3-D and higher-order domains.