The role of signal-to-noise ratio in preserving diagnostic performance and reader confidence level in count-reduced planar nuclear studies undergoing Pixon image processing

Abstract

1432

Objectives: A reduction in acquisition time and/or dose may be clinically valuable in nuclear medicine, resulting in higher throughput, less motion artifact and/or lower radiation exposure. We investigate the role of signal-to-noise ratio (SNR) in the clinical ability of Pixon image processing to preserve image quality in count-reduced planar images by measuring the changes in diagnostic performance and confidence between original images and processed, count-reduced images.

Methods: 12 physicians with little or no previous experience with Pixon processing are presented with 29 typical cases for which image quality is essential for diagnosis obtained at the Geisinger Medical Center (Danville, PA) over a 4-day period (18 bone, 2 Ga, 2 hepatobiliary, 2 V/Q, 3 lymphatic, 2 renal). Count-reduced images are first obtained from the originals using pixel-by-pixel binomial (Poisson preserving) subsampling of 20%, 40% and 60%, and then undergo Pixon image processing. These images are next linearly fused with the unprocessed, count-reduced images, so the resulting SNRs are the same as in the original images. The goal is to test if image quality is preserved by restoring the SNR. The readers are presented with side-by-side pairs of the original and processed images with randomly selected subsampling percentages, case order, and left-right positions. They score the differences in diagnostic performance and reader confidence level between them on a 5-point scale. After correcting for the random left-right positions (unknown to the readers), the possible scores are: -2=original significantly better, -1=original somewhat better, 0=no difference, +1=processed somewhat better, +2=processed significantly better.

Results: [table]

Conclusions: For all subsamplings, the readers, in most cases, see no difference between the original and count-reduced images or prefer the processed ones. For 60% subsampling, the mean score is statistically consistent with 0. For 40% and 20% subsampling, there are statistically significant negative means of less than half a point. For these lower subsamplings, SNR restoration alone is not enough to preserve image quality, and the SNR should be increased.

Distribution of Reader Scores