Evaluating iterative reconstruction performance in computed tomography

Purpose: Iterative reconstruction (IR) offers notable advantages in computed tomography (CT). However, its performance characterization is complicated by its potentially nonlinear behavior, impacting performance in terms of specific tasks. This study aimed to evaluate the performance of IR with both task‐specific and task‐generic strategies. Methods: The performance of IR in CT was mathematically assessed with an observer model that predicted the detection accuracy in terms of the detectability index ( d ′). d ′ was calculated based on the properties of the image noise and resolution, the observer, and the detection task. The characterizations of image noise and resolution were extended to accommodate the nonlinearity of IR. A library of tasks was mathematically modeled at a range of sizes (radius 1–4 mm), contrast levels (10–100 HU), and edge profiles (sharp and soft). Unique d ′ values were calculated for each task with respect to five radiation exposure levels (volume CT dose index, CTDI vol : 3.4–64.8 mGy) and four reconstruction algorithms (filtered backprojection reconstruction, FBP; iterative reconstruction in imaging space, IRIS; and sinogram affirmed iterative reconstruction with strengths of 3 and 5, SAFIRE3 and SAFIRE5; all provided by Siemens Healthcare, Forchheim, Germany). The d ′ values were translated into the areas under the receiver operating characteristic curve (AUC) to represent human observer performance. For each task and reconstruction algorithm, a threshold dose was derived as the minimum dose required to achieve a threshold AUC of 0.9. A task‐specific dose reduction potential of IR was calculated as the difference between the threshold doses for IR and FBP. A task‐generic comparison was further made between IR and FBP in terms of the percent of all tasks yielding an AUC higher than the threshold. Results: IR required less dose than FBP to achieve the threshold AUC. In general, SAFIRE5 showed the most significant dose reduction potentials (11–54 mGy, 77%–84%), followed by SAFIRE3 (7–36 mGy, 50%–61%) and IRIS (6–26 mGy, 37%–50%). The dose reduction potentials highly depended on task size and task contrast, with tasks of lower contrasts and smaller sizes, i.e., more challenging tasks, indicating higher dose reductions. Softer edge profile showed higher dose reduction potentials with SAFIRE3 and SAFIRE5, but not with IRIS. The task‐generic comparison between IR and FBP demonstrated the overall superiority of IR performance, as IR allowed a larger percent of tasks to exceed the threshold AUC: IRIS, 8%–12%; SAFIRE3, 10%–16%; and SAFIRE5, 20%–33%. The improvement with IR was generally more pronounced at lower dose levels. Conclusions: Expanding beyond traditional contrast and noise based assessments of IR, we performed both task‐specific and task‐generic evaluations of IR performance. The task‐specific evaluation demonstrated the dependency of IR's dose reduction potential on task attributes, which can be employed to optimize IR for clinical indications with specific range of size and contrast. The task‐generic evaluation demonstrated IR's overall superiority over FBP in terms of the range of tasks exceeding a threshold performance level, which can be employed for general comparisons between algorithms.