Quantitative metrics of CT images may magnify the potential of radiation dose reduction

This study aimed to explore the relationship between the detectability index (d') and the actual detection performance of small low-contrast lesions and assess the influence of different algorithms and matrices on d'. A phantom was scanned at 9 radiation exposures via a 160-slice CT scanner. All images were reconstructed utilizing filtered back projection, iterative reconstruction, and deep learning (DL) algorithms with both 512 × 512 and 1024 × 1024 matrix. To evaluate the image quality through calculation of the noise power spectrum, task-based transfer function, and d' of the images, the ImQuest software (Duke) was used. A multireader multicase study was conducted to establish the diagnostic accuracy for each set of scan parameters. To compare the relationship between d' and area under the receiver operating characteristic curve, the Spearman rank-sum correlation coefficient was utilized. For DL, the d' values of the 1024-matrix images were lower than that of the 512-matrix images at an ultra-low dose and higher than that of the 512-matrix images at medium and normal doses. The area under the receiver operating characteristic curve of the normal dose 512-filtered back projection images was 0.75. d' is strongly correlated with observer performance (the Spearman rank-sum correlation coefficient = 0.934-0.989). The potential radiation dose reduction according to all quantitative metrics was higher than the true potential (<12-75%). Quantitative metrics may overestimate the potential of iterative reconstruction and DL algorithms to decrease the radiation dose in low-contrast lesions. The d' of different reconstruction methods cannot be directly compared.