Comparison between human and model observer performance in low-contrast detection tasks in CT images: application to images reconstructed with filtered back projection and iterative algorithms.

Abstract

To compare low-contrast detectability (LCDet) performance between a model [non-pre-whitening matched filter with an eye filter (NPWE)] and human observers in CT images reconstructed with filtered back projection (FBP) and iterative [adaptive iterative dose reduction three-dimensional (AIDR 3D; Toshiba Medical Systems, Zoetermeer, Netherlands)] algorithms. Images of the Catphan® phantom (Phantom Laboratories, New York, NY) were acquired with Aquilion ONE™ 320-detector row CT (Toshiba Medical Systems, Tokyo, Japan) at five tube current levels (20-500 mA range) and reconstructed with FBP and AIDR 3D. Samples containing either low-contrast objects (diameters, 2-15 mm) or background were extracted and analysed by the NPWE model and four human observers in a two-alternative forced choice detection task study. Proportion correct (PC) values were obtained for each analysed object and used to compare human and model observer performances. An efficiency factor (η) was calculated to normalize NPWE to human results. Human and NPWE model PC values (normalized by the efficiency, η = 0.44) were highly correlated for the whole dose range. The Pearson's product-moment correlation coefficients (95% confidence interval) between human and NPWE were 0.984 (0.972-0.991) for AIDR 3D and 0.984 (0.971-0.991) for FBP, respectively. Bland-Altman plots based on PC results showed excellent agreement between human and NPWE [mean absolute difference 0.5 ± 0.4%; range of differences (-4.7%, 5.6%)]. The NPWE model observer can predict human performance in LCDet tasks in phantom CT images reconstructed with FBP and AIDR 3D algorithms at different dose levels. Quantitative assessment of LCDet in CT can accurately be performed using software based on a model observer.