Premium
The observer SNR penalty for reconstructions from projections
Author(s) -
Wagner R. F.,
Brown D. G.,
Burgess A. E.,
Hanson K. M.
Publication year - 1984
Publication title -
magnetic resonance in medicine
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.696
H-Index - 225
eISSN - 1522-2594
pISSN - 0740-3194
DOI - 10.1002/mrm.1910010109
Subject(s) - national laboratory , library science , radiological weapon , health science , columbia university , citation , center (category theory) , medicine , physics , computer science , sociology , engineering physics , media studies , medical education , chemistry , radiology , crystallography
Optimization of the efficiency of information collection in medical imaging requires simultaneous consideration of a number of factors, of which some are obvious, others subtle (1, 2). The factors taken singly seem to have little significance, whereas in concert they often lead to possibilities of more than an order of magnitude improvement in imaging efficiency, i.e., signal-to-noise ratio (SNR) squared per unit exposure or per unit exposure time. The purpose of this note is to point out one of the more subtle points related to CT and NMR information collection efficiency that must be considered when the concert is replayed using NM[R themes. We have found that a rigorous science of imaging can be built by studying the performance of the ideal observer of image data (3-6). The performance of real observers of conventional images seems to cluster about a point that falls short of ideal performance by a factor in the neighborhood of 2 when there is sufficient display contrast (7-9). Now, the ideal observer of images containing uncorrelated or white noise tests for the presence of a suspected lesion by a simple weighted average over the area of a lesion; the ideal observer of images containing negative correlations, such as the high-pass or ramp filter noise in CT, must take these correlations into account in testing for the lesion (10). This can be done either by using a rather complicated weighting in the average, or equivalently by first including a step to rewhiten the noise. Our experience with human observers is that they are incapable of this rewhitening step for a range of tasks of signal detection and signal discrimination that we have studied (9). In this work the signal parameters were explicitly specified for the observer a priori. We expect that human observers will continue to lack this capability for more complicated viewing tasks. This inability to cope with the negative correlatiions in image noise leads to what we call the “reconstruction/obsrver penalty” for images reconstructed from projections (“projection reconstruction” or PR images among MMR researchers). In conventional two-dimensional PR images this penalty is rigorously equal to ?r/2 for lesions with a Gaussian profile (3, 4) ; i.e., the exposure or imaging time required to obtain a given SNR is greater by a factor of 7r/2 for real observers of PR images than it would be for real observers of images from data not requiring the reconstruction process, all other factors being equal. In X-ray imaging this point is academic since the only way to avoid the PR process for an axial slice or tomogram is to excise the slice.