Contrast-To-Noise Ratio Is Not An Appropriate Measure of CT Image Quality When Comparing Different Iterative Reconstruction Algorithms
I Reiser*, Z Lu, R Nishikawa, University of Chicago, CHICAGO, ILTH-E-217BCD-1 Thursday 1:00:00 PM - 2:50:00 PM Room: 217BCD
Contrast-to-noise ratio (CNR) is currently one of the pass/fail criteria in acceptance testing and annual quality control of CT scanners as well as for ACR accreditation. However CNR does not take into account background noise correlations. The non-prewhitening observer model signal-to-noise ratio (SNRNPW) on the other hand accounts for noise correlations, and has been shown to correlate with the performance of human observers in the task of detecting low-contrast signals in uniform backgrounds. The purpose of this work was to investigate the differences between these two performance metrics, CNR and SNRNPW.
Images of the ACR CT Accreditation Phantom were acquired on a clinical 64-slice CT scanner. First, phantom images were acquired for four different exposure values and reconstructed with the standard algorithm. Next, phantom images were acquired at one exposure and reconstructed with six additional iterative reconstruction algorithm settings available on the scanner. Slice thickness was 3mm. CNR was computed according to ACR guidelines. For the SNRNPW measurements, 10 image slices were extracted that contained the low-contrast signals, resulting in 40 signal samples for each signal size. SNRNPW was estimated using a template-matching technique.
Increasing exposure from 250mAs to 400mAs improved both CNR and SNRNPW by about 25% and 50%, respectively. However, across the seven reconstruction algorithms, CNR increased by 67%, whereas SNRNPW increased by less than 3%, for the 6-mm, 5-mm and 4-mm diameter signals.
SNRNPW and CNR exhibited similar trends when the exposure was varied. However, for different reconstruction algorithms, the observed increase in CNR was absent in for SNRNPW. This implies that low-contrast signal visibility did not improve, and supports that CNR is not an appropriate figure-of-merit for evaluating dose reduction potential across different reconstruction algorithms.