Quantitative and qualitative evaluation of hybrid iterative reconstruction, with and without noise power spectrum models: A phantom study

Abstract The purpose of this phantom study was to investigate the feasibility of dose reduction with hybrid iterative reconstruction, with and without a noise power spectrum (NPS) model, using both quantitative and qualitative evaluations. Standard dose (SD), three‐quarter dose (TQD), and half‐dose (HD) of radiation were used. Images were reconstructed with filtered back projection (FBP), adaptive iterative dose reduction 3D (AIDR 3D) (MILD, STR), and AIDR 3D enhanced (eAIDR 3D) (eMILD, eSTR). An NPS analysis, task‐based modulation transfer function (MTF task) analysis, and comparisons of low‐contrast detectability and image texture were performed. Although the eAIDR 3D had a higher NPS value in the high‐frequency range and improved image texture and resolution as compared with AIDR 3D at the same radiation dose and iteration levels, it yielded higher noise than AIDR 3D. Additionally, although there was no statistically significant difference between SD‐FBP and the TQD series in the comparison of the mean area under the curve (AUC), the mean AUC was statistically significantly different between SD‐FBP and the HD series. NPS values in the high‐frequency range, 10% MTF task values, low‐contrast detectability, and image textures of TQD‐eMILD were comparable to those of SD‐FBP. Our findings suggested that using eMILD can reduce the radiation dose by 25%, while potentially maintaining diagnostic performance, spatial resolution, and image texture; this could support selecting the appropriate protocol in a clinical setting.

that can allow a reduction in radiation dose. 2-8 Although these techniques iteratively reduce noise in the image space, raw data, or both, IR techniques have also been reported to produce changes in image texture. 5,[9][10][11][12] Adaptive iterative dose reduction 3D (AIDR 3D) (Toshiba Medical Systems, Otawara, Japan) is a hybrid IR technique that uses a scanner model and statistical noise model, together with projection noise estimation in the raw data domain, to reduce photon and electronic noise. 13,14 Several previous studies have indicated that AIDR 3D improves image quality and reduces dose in a manner comparable to IR. 15,16 In contrast, it has also been reported that resolution changes in accordance with radiation dose, iterative strength, and contrast, and that low contrast detectability is not necessarily improved at low dose levels. 17,18 AIDR 3D Enhanced (eAIDR 3D) is an IR-mounted noise power spectrum (NPS) model that preserves high-frequency noise in the NPS and is expected to offer improved image texture and resolution as compared to AIDR 3D. Yet, to the best of our knowledge, no study has yet evaluated the image quality characteristics of eAIDR 3D in detail. Additionally, it is known that quantitative evaluations of IR, such as contrast-to-noise ratio analyses, diverge from qualitative evaluations, because IR is a nonlinear reconstruction method. 18 Therefore, both quantitative and qualitative evaluations are necessary for assessing IR image quality.
The purpose of this study was to investigate the feasibility of dose reduction with hybrid iterative reconstruction, with and without an NPS model, in a phantom using both quantitative and qualitative evaluations.

| MATERIALS AND METHODS
The NPS analysis and task-based modulation transfer function task (MTF task ) analysis were performed as quantitative evaluations.
Low-contrast detectability was compared using a receiver operating characteristic (ROC) curve analysis and visual image texture was compared using Scheffe's method of paired comparisons, as qualitative evaluations.
As described below, the phantom that was used for quantitative evaluation was filled with diluted contrast medium. Therefore, AEC for quantitative evaluation indicated a larger dose level than qualitative evaluation, and consequently, two different dose levels were used for qualitative and quantitative evaluations.
For quantitative evaluations, non-helical scanning with an 80 9 0.5 mm 2 detector configuration was performed to eliminate the influence of table movement on the image-averaging process, as described later. 17 For qualitative evaluation, helical scanning (pitch factor, 0.844) with a 32 9 1.0 mm 2 detector configuration was used, assuming clinical settings. All images were reconstructed with a dis-

2.B | Noise power spectrum (quantitative) analysis
An acrylic phantom with a diameter of 200 mm was filled with water and used for NPS analysis [ Fig. 1(a)]. The acrylic phantom was placed at the isocentre of the CT scanner. To acquire NPS for each reconstructed image, a region of interest (ROI) of 100 cm 2 (256 9 256 pixels) was placed on the centre of the image, as shown in Fig. 1 Fig. 1(b)]. The acrylic phantom was placed at the isocenter of the CT scanner. To acquire low-noise images for the MTF task analysis, 100 or more scans were performed with the same table position for each radiation dose level and these were averaged for each protocol using image-averaging techniques. 17 Furthermore, images were reacquired after the liquid in the phantom was adjusted with diluted contrast medium (60 HU) to obtain a contrast of 10 HU between background and soft tissue (70 HU). To acquire MTF task values for each averaged image, an ROI of 9.77 cm 2 (80 9 80 pixels) was placed around the three objects as shown in Fig. 1   T A B L E 2 Steel-Dwass test of relative noise for each protocol.   HD-eMILD were similar to that of SD-FBP. Furthermore, the eAIDR 3D had a lower NPS value in the low frequency range and had a higher NPS value in the high frequency range than the AIDR 3D, at the same radiation dose and iteration level. The relative noise value of eAIDR 3D was higher than that of AIDR 3D at the same radiation dose level and iteration level (Tables 1 and 2). Table 3 and Fig. 3 show the 10% MTF task values and MTF task curves.

3.B | Modulation transfer function task analysis
Although both AIDR 3D and eAIDR 3D had lower 10% MTF task values with lower CT values, lower radiation doses, and higher iteration levels, the 10% MTF task values were higher for eAIDR 3D than for AIDR 3D at the same radiation doses and iteration levels. The 10% MTF task values of all AIDR 3D protocols were equal to or lower than those of SD-FBP. In contrast, the 10% MTF task values of eAIDR 3D at TQD and HD tended to be higher at 120 and 300 HU, and tended to be equal to or lower than those of SD-FBP at 10 and 70 HU. The 10% MTF task values of TQD-eMILD at 10 and 70 HU were equal to those of SD-FBP.

| DISCUSSION
Our study demonstrated that eMILD allowed a 25% reduction in radiation dose while maintaining diagnostic performance, spatial resolution, and image texture. Image quality has not previously been compared between AIDR 3D and eAIDR 3D using both quantitative and qualitative evaluations. We found that NPS values in the high frequency range, 10% MTF task values, low-contrast detectability, and image texture of TQD-eAIDR 3D were superior to those of TQD-AIDR 3D, and similar to those of SD-FBP. These findings are important because they can guide protocol selection in a clinical setting.
T A B L E 4 Average AUC and 95% CI for each protocol. Solomon et al. reported that it is possible to use an NPS to compare image texture quantitatively. 40 Our NPS analysis findings that the nNPS curves of TQD-eMILD and HD-eMILD were close to that of SD-FBP indicated that the image texture of TQD-eMILD and HD-eMILD is similar to that of SD-FBP.
Our findings that 10% MTF task values changed in accordance with contrast are consistent with a previous study by Richard et al., and represent a feature of nonlinear processing in IR. 25 We found that although eAIDR 3D improved 10% MTF task values at 120 HU or higher, it did not improve 10% MTF task values at 70 HU or less.
Therefore, in clinical settings, eAIDR 3D may be useful for enhanced CT or CT angiography (i.e., imaging with high contrast levels).
We found that the mean AUC value for low-contrast detectability of TQD-eMILD was significantly higher than that of TQD-MILD and was comparable to that of SD-FBP, and that image texture in TQD-eMILD was similar to that in SD-FBP. Thus, we suggest that TQD-eMILD is desirable for maintaining both diagnostic performance and image texture, while reducing the dose of radiation required. In contrast, the mean AUCs for the low-contrast detectability values of HD-MILD, HD-STR, HD-eMILD, and HD-eSTR were significantly lower than that of SD-FBP, such that a 50% reduction in radiation dose by AIDR 3D or eAIDR 3D may not be feasible for the detection of small low-contrast lesions.

| CONCLUSION
We suggest that the use of eMILD can facilitate a 25% reduction in radiation dose while potentially maintaining diagnostic performance, spatial resolution, and image texture.

ACKNOWLEDG MENTS
The authors acknowledge Syota Masuda, Toshihiko Machida, Shigeki Miura, and Masahiro Kozaki for performing qualitative evaluations and Masakazu Hasegawa, Toshio Watanabe, and Sou Tsushima for technical assistance.

Masahiro Jinzaki received a research grant from Toshiba Medical
Systems. The remaining authors have no financial disclosures to make in relation to this study.

SOURCES OF FUN DING
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.