Detectability of Small Low-Attenuation Lesions With Deep Learning CT Image Reconstruction: A 24-Reader Phantom Study.
Journal:
AJR. American journal of roentgenology
Published Date:
Sep 21, 2022
Abstract
Iterative reconstruction (IR) techniques are susceptible to contrast-dependent spatial resolution, limiting overall radiation dose reduction potential. Deep learning image reconstruction (DLIR) may mitigate this limitation. The purpose of our study was to evaluate low-contrast detectability performance and radiation-saving potential of a DLIR algorithm in comparison with filtered back projection (FBP) and IR using a human multireader noninferiority study design and task-based observer modeling. A dual-phantom construct, consisting of a low-contrast detectability module (21 low-contrast hypoattenuating objects in seven sizes [2.4-10.0 mm] and three contrast levels [-15, -10, -5 HU] embedded within liver-equivalent background) and a phantom, was imaged at five radiation exposures (CTDI range, 1.4-14.0 mGy; size-specific dose estimate, 2.5-25.0 mGy; 90%-, 70%-, 50%-, and 30%-reduced radiation levels and full radiation level) using an MDCT scanner. Images were reconstructed using FBP, hybrid IR (ASiR-V), and DLIR (TrueFidelity). Twenty-four readers of varying experience levels evaluated images using a two-alternative forced choice. A task-based observer model (detectability index []) was calculated. Reader performance was estimated by calculating the AUC using a noninferiority method. Compared with FBP and IR methods at routine radiation levels, DLIR medium and DLIR high settings showed noninferior performance through a 90% radiation reduction (except DLIR medium setting at 70% reduced level). The IR method was non-inferior to routine radiation FBP only for 30% and 50% radiation reductions. No significant difference in was observed between routine radiation FBP and DLIR high setting through a 70% radiation reduction. Reader experience was not correlated with diagnostic accuracy ( = 0.005). Compared with FBP or IR methods at routine radiation levels, certain DLIR algorithm weightings yielded noninferior low-contrast detectability with radiation reductions of up to 90% as measured by 24 human readers and up to 70% as assessed by a task-based observer model. DLIR has substantial potential to preserve contrast-dependent spatial resolution for the detection of hypoattenuating lesions at decreased radiation levels in a phantom model, addressing a major shortcoming of current IR techniques.