Low contrast detectability performance of model observers based on CT phantom images: kVp influence.

Abstract This paper studies low contrast detectability (LCD) performance of two model observers in CT phantom images acquired at different kVp levels and compares the results with humans in a 2-alternative forced choice experiment (2-AFC). Images of the Catphan phantom with objects of different contrasts (0.5 and 1%) and diameters (2–15 mm) were acquired in an Aquilion ONE 320-detector row CT (Toshiba Medical Systems, Tokyo, Japan), in two experiments, selecting (80–100–120–135 kV) with fixed mAs and varying the mAs to keep the dose constant, respectively. Four human observers evaluated the objects visibility obtaining a proportion correct (PC) for each case. LCD was also analyzed with two model observers (non-prewhitening matched filter with an eye filter, NPWE, and channelized Hotelling observer with Gabor channels, CHO). Object contrast was affected by kV, with differences up to 17% between the lowest and highest kV. Both models overestimated human performance and were corrected by efficiency and internal noise factors. The NPWE model reproduced better the human PC values trends showing Pearson's correlation coefficients ≥0.976 (0.954–0.987, 95% CI) for both experiments, whereas for CHO they were ≥0.706 (0.493–0.839). Bland–Altman plots showed better agreement between NPWE and humans being the average difference Δ and the range of the differences Δ±2σ (σ, standard deviation) of Δ=−0.3%, Δ±2σ = [−4.0%,4.5%]. For CHO, Δ=−1.2%, Δ± 2σ= [−10.7%,8.3%]. The NPWE model can be a useful tool to predict human performance in CT low contrast detection tasks in a standard phantom and be potentially used in protocol optimization based on kV selection.