A feasibility study of realizing low-dose abdominal CT using deep learning image reconstruction algorithm

Abstract

OBJECTIVES:

To explore the feasibility of achieving diagnostic images in low-dose abdominal CT using a Deep Learning Image Reconstruction (DLIR) algorithm.

METHODS:

Prospectively enrolled 47 patients requiring contrast-enhanced abdominal CT scans. The late-arterial phase scan was added and acquired using lower-dose mode (tube current range, 175–545 mA; 80 kVp for patients with BMI ≤24 kg/m² and 100 kVp for patients with BMI > 24 kg/m²) and reconstructed with DLIR at medium setting (DLIR-M) and high setting (DLIR-H), ASIR-V at 0% (FBP), 40% and 80% strength. Both the quantitative measurement and qualitative analysis of the five types of reconstruction methods were compared. In addition, radiation dose and image quality between the early-arterial phase ASIR-V images using standard-dose and the late-arterial phase DLIR images using low-dose were compared.

RESULTS:

For the late-arterial phase, all five reconstructions had similar CT value (P > 0.05). DLIR-H, DLIR-M and ASIR-V80% images significantly reduced the image noise and improved the image contrast noise ratio, compared with the standard ASIR-V40% images (P < 0.05). ASIR-V80% images had undesirable image characteristics with obvious “waxy” artifacts, while DLIR-H images maintained high spatial resolution and had the highest subjective image quality. Compared with the early-arterial scans, the late-arterial phase scans significantly reduced the radiation dose (P < 0.05), while the DLIR-H images exhibited lower image noise and good display of the specific image details of lesions.

CONCLUSIONS:

DLIR algorithm improves image quality under low-dose scan condition and may be used to reduce the radiation dose without adversely affecting the image quality.

Keywords

DLIR algorithm low-dose abdominal CT image quality

1 Introduction

It has been more than 120 years since Roentgen discovered X-rays in 1895. CT, as a non-invasive examination method, has been widely used in clinical practice due to its advantages of high resolution, fast imaging speed and easy operation. In fact, almost a third of CT examinations involve the abdomen, and abdominal and pelvic CT has become a key player in the diagnosis and management of patients with acute abdominal pain [1]. However, the radiation dose and potential carcinogenic effect of CT in clinical application have been of great concern [2]. Automated tube current modulation (ATCM), optimized x-ray tube voltage, and better use of iterative image reconstruction have allowed the maintenance of good CT image quality with reduced radiation dose [3, 4]. Image noise is an important determinant of CT image quality, which is inversely proportional to X-ray dose. The reduction of tube voltage or tube current will lead to the reduction of radiation dose, which is also related to the increase of image noise, which means that the reduction of radiation dose may at the same time degrade the image quality to various degrees [5]. Therefore, how to maintain image quality while reducing radiation dose has become a research hotspot.

Some CT manufacturers have developed a new generation of CT image reconstruction based on deep learning. The deep learning image reconstruction (DLIR) algorithms developed by GE Healthcare and Canon Medical system feature a deep neural network (DNN), which is trained respectively with high-quality filtered back projection (FBP; TrueFidelity™, GE Healthcare) and MBIR (Advanced intelligent Clear-IQ Engine (AiCE), Canon Medical System) datasets to learn how to differentiate noise from signals [6]. DLIR is the latest image reconstruction algorithm developed by GE to provide better images under low radiation dose in real time. In this technique, deep convolutional neural network–based models are used to emulate the image textures of the standard-dose filtered back projection (FBP) reconstructions while significantly reducing image noise and streak artifacts, improving low contrast lesion detectability, and high resolution. This vendor-specific version of DLIR currently allows the selection of low, medium, and high reconstruction strengths, which translates into the degree of desired noise reduction [7]. The purpose of this study was to explore the feasibility of achieving diagnostic images in low-dose contrast-enhanced CT with DLIR using qualitative reader evaluations and quantitative region-of-interest (ROI) measurements in comparison with our start-of-the-art ASIR-V images at the standard 40% strength (40% ASIR-V).

2 Materials and methods

This prospective study enrolled 47 patients who were required to undergo contrast-enhanced upper abdominal CT examination from May to July 2020. All patients provided informed consent, and this study was approved by the Ethics Committee of the hospital. Precautions were taken before contrast-enhanced scans: (1) Inform patients of the purpose and significance of contrast-enhanced scans to obtain the patient’s cooperation. (2) Ask patients in detail about his/her history and whether he/she has a history of allergy. If he/she has a history of iodine allergy, severe cardiac, liver and renal insufficiency, other important organ dysfunction, abnormal coagulation mechanism, etc., or has asthma, diabetes, hyperthyroidism, etc, he/she should be excluded from the study or carefully examined before enrollment. (3) Clarify the adverse reactions that may occur when injecting contrast agent to patients and sign the informed consent for using contrast agent.

2.1 Clinical case data collection and CT scanning methods

Patient demographic information (age, sex, height, weight, and body mass index) were recorded. The study group consisted of 28 males and 19 females with a mean±SD age of 54.55±11.53years (range, 24–77years), an average patient weight of 64.06±11.50 kg (range, 41–88 kg), and mean body mass index of 23.22±3.21 kg/m² (range, 16.02–29.40 kg/m²).

All patients underwent multi-phase contrast-enhanced CT (early-arterial, late-arterial, portal and delay phases) of the abdomen consecutively on a 256-slice CT system (Revolution CT, GE Healthcare). The late-arterial phase acquisition was added in the scan protocol using a lower-dose protocol as follows: gantry rotation speed, 0.6 second; pitch, 0.5:1; table speed, 40 mm/rotation; beam collimation, 80 mm; detector configuration, 128×0.625 mm; tube current modulation range, 175–545 mA; 80 kVp for patients with BMI ≤24 kg/m² and 100 kVp for patients with BMI > 24 kg/m²; noise index, 12 based on 5 mm reconstruction thickness (40% ASIR-V). The other three phases were scanned using the standard-dose for routine clinical diagnosis as follows: gantry speed, 0.6 second; pitch, 0.5:1; table speed, 40 mm/rotation; beam collimation, 80 mm; detector configuration, 128×0.625 mm; tube current modulation range, 175–545 mA; 120 kVp; noise index, 8 based on 5 mm reconstruction thickness (40% ASIR-V).

Weight-based IV contrast injection was used with settings of 0.6 g I/kg allowing a range of 80–100 mL and a fixed injection duration of 6 seconds. Iohexol (Omnipaque 350, GE Healthcare) was used as the IV contrast agent. The bolus tracking technique was used with the trigger threshold set at 150 HU in the abdominal aorta at the level of the celiac artery. The early-arterial phase scan started with a scan delay of 6 seconds after triggering and the late-arterial phase scan started with 3.2 seconds delay after the early-arterial phase acquisition. The portal phase images were obtained with a delay of 25 seconds and the delay phase images were obtained with a delay of 60 seconds after triggering.

Five sets of axial reconstructions were generated using the low-dose scan data in the late-arterial phase for each patient with the standard reconstruction kernel at 1.25 mm slice thickness and 1.25 mm slice interval: 0% ASIR-V(FBP), ASIR-V with 40% strength (40% ASIR-V), ASIR-V with 80% strength (80% ASIR-V), and DLIR with the medium-strength (DLIR-M) and high-strength (DLIR-H). For other phases with the standard-dose, images were reconstructed using only the standard 40% ASIR-V algorithm for diagnosis. The standard-dose, early-arterial phase 40% ASIR-V images were also used as the reference standard for comparing the DLIR low-dose images in the late-arterial phase.

2.2 Qualitative analysis

The reconstructed axial images were reviewed under standard clinical protocols with high-resolution monitors at a two-monitor PACS workstation. The order of the reconstructions was randomized for displaying, and all identifying patient information and annotations were removed. The reconstructions were linked so that identical anatomic levels could be evaluated during scrolling. Images were presented with window settings of width 400 HU and level 40 HU. The readers were allowed to scroll and change the window settings and zoom while reviewing the cases.

Two radiologists blinded to scanning and reconstruction details independently reviewed each case after receiving standardized instructions. An electronic copy of the instructions was provided during the reading sessions for reference. The readers were fellowship trained in abdominal imaging and had 5 and 10 years of additional experience beyond fellow-ship in reading abdominal CT images. There was no time limit for review. Two radiologists rated the overall images quality and lesion diagnostic confidence of the five types of reconstructions in the late-arterial phase and the 40% ASIR-V images in the early-arterial phase using a 5-point Likert scale: 5, excellent; 4, above average; 3, average; 2, below average; 1, poor.

The following method was used for evaluating lesions for the two readers: the first reader randomly assigned each case was instructed to annotate and save marks on all solid organ lesions defined as any noncalcified, focal, solid organ abnormality larger than 2 mm of either increased or decreased relative attenuation [8]. The second reader was directed to evaluate the appearance of any lesions present on the marked key images to ensure that both readers were evaluating the same lesions [7]. As many of the patients had the diagnosis of lymphoma, lymphadenopathy was also counted as a lesion. If there were multiple lesions of the same type (e.g., liver cysts or enlarged lymph nodes), a single representative lesion was chosen and flagged. The type of lesion present on the flagged images was recorded.

2.3 Qualitative analysis

The image noise was indicated by the standard deviation (SD) of the CT numbers over a region of interest (ROI) in a homogeneous substance. A radiologist who was not involved in grading the examinations measured CT number value and the image noise by placing a 1 cm³ ROI in 5 standard locations: paraspinal musculature, liver parenchyma, aorta, spleen and renal cortex. To ensure consistency, ROI placement per patient was matched as closely as possible in the CT series (as shown in Fig. 1).

Fig. 1

Low-dose CT images with 5 different reconstructions in the late-arterial phase for a 56-year-old woman with body mass index of 23.9 kg/m². A, FBP; B, 40% ASIR-V; C, 80% ASIR-V; D, DLIR-M; E, DLIR-H. ROI measurements are also shown.

For each reconstruction, the contrast-to-noise ratio (CNR) relative to muscle was calculated for the liver parenchyma, aorta, spleen and renal cortex as (ROIi –ROIm) / SD, where ROIi is the attenuation for the anatomy of interest, ROIm is the attenuation of paraspinal muscles, and SD is the image noise of paraspinal muscles.

2.4 Statistical analysis

Summary statistics for image quality scores, attenuation, SD, CNR were calculated as means, SDs, and ranges for each item and series. Matched ordinal scale variables (for example image quality scores in the late-arterial phase) were compared with Friedman’s test, independent ordinal scale variables with Kruskal-Wallis One-way ANOVA test and interval scale variables with ANOVA test. Histograms were used to show the distributions of Likert scale reader scores across reconstructions. Statistical significance was set at P < 0.05. The weighted Cohen kappa statistic was used to assess agreement between the two readers.

3 Result

3.1 Subjective image quality

The images of the 5 different reconstruction methods with ROI selection under the low-dose scanning in the late-arterial phase are shown in Fig. 1. On the 5-point Likert scale for the overall late-arterial phase low-dose abdomen image quality, the readers rated DLIR-H as the best. The differences in the ratings of these 5 reconstructions were statistically significant. The 80% ASIR-V series was often described as having an overly smoothed appearance and was deemed to be not acceptable for diagnosis in many cases (as shown in Fig. 2). In order to avoid excessive data volume, for lesion evaluation, we only presented those from the 40% ASIR-V, DLIR-M and DLIR-H images in the late-arterial phase and compared them with the 40% ASIR-V images in the early-arterial phase with the standard dose. The results are shown in Fig. 3. Except for one case of multiple liver metastases, two cases of multiple liver cyst, one case of frequent kidney cyst, one case of multiple lesions of liver cancer after interventional, two cases of retroperitoneal multiple enlarged lymph nodes, one case of left kidney seeper, two cases no significant lesions, the readers identified a total of 140 lesions in this experiment including five liver cancer, one duodenal cancer, 20 liver cyst, 40 renal cyst, two adrenal adenoma, one adrenal cyst, one intraductal papillary mucinous tumor (IPMN) of the pancreas, 43 liver hemangioma, 16 after the intervention of liver cancer lesions, 11 liver metastases. Similarly, for lesion evaluation, DLIR-H had the highest score. Kappa test showed good consistencies (Kappa > 0.7) between the two readers. The lesions subject score comparison between the early-arterial phase and late-arterial phase with different doses demonstrated that DLIR-H achieved similar image quality as the standard 40% ASIR-V images with 57% dose reduction (as shown in Table 1, Figs. 4 and 5).

Fig. 2

Histogram of reader subjective quality scores for overall abdomen image quality with five different reconstructions in the low-dose late-arterial phase. Note that due to the use of lower dose and thinner slice thickness reconstruction than the standard settings, the FBP image (A) in the late-arterial phase was not suited for clinical diagnosis. On the other hand, the DLIR images (D and E) were fully satisfactory for the diagnostic purpose.

Fig. 3

Histogram of reader subjective scores for lesion conspicuity with 3 different reconstructions in both the low-dose of late-arterial phase and standard-dose of early-arterial phase.

Table 1

Radiation dose comparison

	CTDIvol (mGy)	DLP (mGy·cm)	ED (mSv)
Early-arterial phase	20.29±0.86	575.75±42.84	8.64±0.64
Late-arterial phase	8.62±3.31	247.23±101.43	3.71±1.52
Z	–8.355	–8.352	–8.352
P	0.000	0.000	0.000

Fig. 4

Images of a 56-year-old woman with body mass index of 21.1 kg/m². A, image in the early-arterial phase with standard radiation dose, A small cyst lesion could be seen in the right lobe of the liver; B, 40% ASIR-V image in the late-arterial phase with 57% reduced radiation dose; C: DLIR-M image in the late-arterial phase; D: DLIR-H image in the late-arterial phase. The DLIR-H image showed much better noise performance than the low-dose 40% ASIR-V image. In addition, the DLIR-H image demonstrated similar spatial resolution, lesion conspicuity for the cystic lesion and lower image noise even with 57% reduced dose compared with the image in the early-arterial phase.

Fig. 5

Images of a 58-year-old man with body mass index of 17.3 kg/m² and with liver metastasis from colon cancer. A, early arterial, three metastatic foci can be seen on this layer, with poor boundaries, among which the lesion indicated by the blue arrow is small; B, 40% ASIR-V image in the late-arterial phase; C: DLIR-M image in the late-arterial phase; D: DLIR-H image in the late-arterial phase; DLIR-H can show the metatheses well.

3.2 Objective image quality

Objective image quality results showed that there was no difference in CT values under different reconstruction methods. SD values were significantly reduced and CNR significantly improved under DLIR-H and 80% ASIR-V, compared with other reconstruction methods (as shown in Table 2 and Table 3).

Table 2
Mean attenuation, noise in the abdomen

Paraspinal musculature Liver parenchyma Aorta Spleen Renal cortex

Attenuation Nosie Attenuation Nosie Attenuation Nosie Attenuation Nosie Attenuation Nosie

(HU) (HU) (HU) (HU) (HU) (HU) (HU) (HU) (HU) (HU)

DLIR-M 62.31±7.87 14.27±2.15 82.36±14.45 14.19±2.42 360.24±99.01 17.93±3.24 152.90±33.66 16.03±4.43 208.95±44.60 19.81±5.94

DLIR-H 62.51±8.97 9.72±1.32 82.41±14.87 10.01±1.66 359.97±98.24 11.96±2.14 154.05±33.57 12.28±4.22 207.82±43.75 16.67±6.28

FBP 63.13±8.79 26.00±2.86 82.39±15.49 25.93±3.10 359.49±97.87 34.59±4.23 152.71±33.77 26.11±4.44 206.58±46.81 31.66±4.10

40% ASIR-V 62.51±7.42 17.70±2.26 82.22±15.43 18.04±2.38 359.50±98.16 24.81±3.94 152.80±33.97 18.59±3.91 208.24±46.05 23.05±4.01

80% ASIR-V 62.85±7.52 10.19±1.74 81.84±14.98 9.27±1.68 359.20±98.15 15.31±4.13 153.22±34.41 12.12±4.85 206.79±45.49 15.02±4.19

F 0.074 175.34 0.012 176.96 0.001 174.887 0.012 145.209 0.023 140.554

P 0.990 0.000 1.000 0.000 1.000 0.000 1.000 0.000 0.999 0.000

	Paraspinal musculature	Liver parenchyma	Aorta	Spleen	Renal cortex
DLIR-M	62.31±7.87	14.27±2.15	82.36±14.45	14.19±2.42	360.24±99.01	17.93±3.24	152.90±33.66	16.03±4.43	208.95±44.60	19.81±5.94
DLIR-H	62.51±8.97	9.72±1.32	82.41±14.87	10.01±1.66	359.97±98.24	11.96±2.14	154.05±33.57	12.28±4.22	207.82±43.75	16.67±6.28
FBP	63.13±8.79	26.00±2.86	82.39±15.49	25.93±3.10	359.49±97.87	34.59±4.23	152.71±33.77	26.11±4.44	206.58±46.81	31.66±4.10
40% ASIR-V	62.51±7.42	17.70±2.26	82.22±15.43	18.04±2.38	359.50±98.16	24.81±3.94	152.80±33.97	18.59±3.91	208.24±46.05	23.05±4.01
80% ASIR-V	62.85±7.52	10.19±1.74	81.84±14.98	9.27±1.68	359.20±98.15	15.31±4.13	153.22±34.41	12.12±4.85	206.79±45.49	15.02±4.19
F	0.074	175.34	0.012	176.96	0.001	174.887	0.012	145.209	0.023	140.554
P	0.990	0.000	1.000	0.000	1.000	0.000	1.000	0.000	0.999	0.000

Table 3

Mean contrast-to-noise ratio in the abdomen

	Liver CNR value	Aorta CNR value	Spleen CNR value	Renal cortex CNR value
DLIR-M	1.41±0.92	16.79±5.27	5.83±2.25	7.68±2.27
DLIR-H	2.00±1.57	25.04±7.62	7.59±2.28	9.34±2.91
FBP	0.74±0.56	8.63±3.01	3.50±1.36	4.51±1.31
40% ASIR-V	1.10±0.81	12.05±4.16	4.92±1.76	6.35±1.77
80% ASIR-V	2.05±1.50	19.92±7.11	8.02±3.38	9.86±2.77
F	114.468	175.47	137.804	135.898
P	0.000	0.000	0.000	0.000

4 Discussion

The FBP reconstruction algorithm is a traditional CT image reconstruction algorithm, which has been widely used in clinic due to its advantages of high speed and convenience [9]. Conventional FBP technique is associated with greater image noise and artifacts in low signal conditions because it does not take into account photon statistics and specific hardware details, such as focal spot, detectors size, and shape [10]. FBP was the method of choice for decades, until the first iterative reconstruction (IR) technique was clinically introduced in 2009 [11]. Some studies have reported that IR have reduced low contrast lesion detectability and spatial resolution compared with FBP methods [12]. Adaptive statistical iterative reconstruction (ASIR) has allowed image reconstruction on the same CT data to significantly reduce image noise compared to FBP [13]. The radiologist can specify the amount of ASIR incorporated into the final image by choosing a percentage from 10% to 100%. Mathematically, the ASIR percentage is simply a linear combination of the original FBP image and the full 100% ASIR image. In general, as the percentage of ASIR increases, the image noise decreases [14]. The Model-based Iterative Reconstruction (MBIR) represents a more advanced and complex version of the IR. MBIR algorithms include Veo, advanced model-based iterative reconstruction, iterative model reconstruction, and forward-projected model-based iterative reconstruction solution [15]. MBIR provides stable noise reduction, but takes a long time to rebuild, and has been criticized for significantly changing the texture information of the image [16]. Veo as the earliest GE Healthcare introduced the second generation of IR algorithm, calculation process is complex and the reconstruction time is relatively long [17, 18]. Adaptive statistical iterative reconstruction-V (ASIR-V) is the latest generation of commercial IR developed by GE. It has the potential for clinically feasible dose reduction with better image quality than conventional ASIR and a shorter imaging processing time than MBIR [19]. ASIR-V combined with various iterative reconstruction algorithms has the advantages of noise reduction, artifact elimination, radiation dose reduction, contrast enhancement and fast calculation speed [16, 20]. However, concerns remain regarding limitations in task-specific diagnoses, such as liver lesion detection, when radiation dose levels are reduced [21]. Recently, DLIR methods using deep convolutional neural networks have been proposed to enable dose reduction while maintaining the diagnostic performance of CT [22 –25].

In this study we explored the feasibility of achieving diagnostic images in low-dose abdominal CT using a Deep Learning Image Reconstruction (DLIR) algorithm. Our preliminary study found that DLIR could significantly reduce image noise while maintaining spatial resolution and favorable noise texture. DLIR displayed robust noise reduction, resulting in significantly improved CNR measured in the liver, spleen, aorta and renal cortex. DLIR improved the image quality of contrast-enhanced CT scans of the abdomen compared to the standard 40% ASIR-V images with significantly better scores in overall image quality, lesion diagnostic confidence and image noise and image texture.

There have been some preliminary studies on DLIR. Phantom study showed that DLIR improves the image quality of CT scans, especially those performed at medium- or reduced radiation doses [24]. And DLIR reduced noise magnitude and improved spatial resolution and detectability without changing noise texture relative to FBP. Images obtained with DLIR seem to indicate more potential for dose optimization process than those obtained with conventional IR algorithms, with reconstruction time adapted to clinical use [6]. At submillisievert chest and abdominopelvic CT doses, DLIR enables image quality and lesion detection superior to commercial IR and FBP images. The image quality improvements with DLIR relative to both FBP and IR techniques should allow users to reduce radiation dose when using DLIR in chest and abdominal CT examinations [26]. Study has demonstrated that DLIR significantly reduces the image noise in low-dose chest LDCT scan images compared with ASIR-V algorithm 30% while maintaining superior image quality [27]. Benz DC et al. showed that DLIR significantly reduces noise in CCTA compared to ASIR-V, yielding superior image quality at equal diagnostic accuracy [28].

The implications of our study included favorable results of image quality and texture associated with DLIR-H in abdominal images compared with FBP 40% ASIR-V 80% ASIR-V and DLIR-M images. Although 80% ASIR-V also significantly reduced noise and improved CNR, the images were overly smooth which negatively affected lesion diagnosis. Reader scores and comments during the study generally favored high-strength DLIR, which produced the lowest degree of image noise. DLIR-H enabled image noise reduction and significantly improved the objective and subjective image qualities without over smoothing, this is similar to the results of other studies [29, 30]. We found that compared with the images in the early-arterial phase with standard dose, the images reconstructed with DLIR-H at the late-arterial phase with only 43% dose were also very good at showing the lesions, and were no-inferior in terms of overall image noise and lesion conspicuity (Fig. 4). The prospective study by Jensen CT et al indicated that CT evaluation of small low-contrast liver lesions is compromised in the setting of modest radiation dose reduction and that iterative reconstructions could not maintain observer performance [21]. Among the cases we collected in our study, there were also some cases of small liver metastasis, and we found that DLIR-H could clearly showed these smaller metastatic foci under low-dose scanning (Fig. 5). This demonstrated that DLIR-H could maintain the diagnosis for lesions, especially small lesions in reduced radiation dose scanning, which is beneficial in practical clinical applications. This is similar to the results of the previous study evaluating the performance of DLIR algorithm [23].

There are limitations to our study. It had a small sample size, and we did not perform a power analysis to determine the minimum number of patients required. In order to evaluate the same lesions of the same patient with different radiation doses and reconstruction algorithms while minimizing the radiation dose received by the patients, we compared image noise and lesion evaluation of the low-dose and standard-dose scans acquired at slightly different enhancement phases. Even through the image noise comparison should not be impacted by the difference of phases, there will be slight changes in lesion enhancement degree and appearance, which may have a certain influence on the results.

5 Conclusions

In conclusion, DLIR-H significantly reduces image noise and improves overall image quality of low-dose abdominal CT examinations and provides similar lesion diagnostic ability with 57% radiation dose reduction compared with the state-of-the-art 40% ASIR-V algorithm.

Conflict of interest and statement

None of the authors have any conflict of interest.

The article does not contain information about medical device(s)/drug(s).

References

Stoker

, van Randen

, Lameris

and Boermeester

M.A.

, Imaging patients with acute abdominal pain, Radiology 253(1) (2009), 31–46.

Coakley

F.V.

, Gould

, Yeh

B.M.

and Arenson

R.L.

, CT radiation dose: What can you do right now in your practice? American Journal of Roentgenology 196(3) (2011), 619–625.

Kaza

R.K.

, Platt

J.F.

, Goodsitt

M.M.

, et al., Emerging techniques for dose optimization in abdominal CT, Radiographics 34(1) (2014), 4–17.

Harun

H.H.

, Karim

M.K.A.

, Abbas

, et al., Effect of iterative reconstruction algorithm levels on noise index and figure-of-merit in CT pulmonary angiography examinations, Journal of X-ray Science and Technology 28(5) (2020), 893–903.

Kalra

M.K.

, Maher

M.M.

, Toth

T.L.

, et al., Strategies for CT radiation dose optimization, Radiology 230(3) (2004), 619–628.

Greffier

, Hamard

, Pereira

, et al., Image quality and dose reduction opportunity of deep learning image reconstruction algorithm for CT: a phantom study, Eur Radiol 30(7) (2020), 3951–3959.

Jensen

C.T.

, Liu

, Tamm

E.P.

, et al., Image quality assessment of abdominal CT by use of new deep learning image reconstruction: initial experience, American Journal of Roentgenology 215(1) (2020), 50–57.

Pickhardt

P.J.

, Lubner

M.G.

, Kim

D.H.

, et al., Abdominal CT with model-based iterative reconstruction (MBIR): initial results of a prospective trial comparing ultralow-dose with standard-dose imaging, American Journal of Roentgenology 199(6) (2012), 1266–1274.

Pan

, Sidky

E.Y.

and Vannier

, Why do commercial CT scanners still employ traditional, filtered back-projection for image reconstruction? Inverse Problems 25(12) (2009), 123009.

10.

Singh

, Kalra

M.K.

, Hsieh

, et al., Abdominal CT: comparison of adaptive statistical iterative and filtered back projection reconstruction techniques, Radiology 257(2) (2010), 373–383.

11.

Willemink

M.J.

and Noël

P.B.

, The evolution of image reconstruction for CT—from filtered back projection to artificial intelligence, European Radiology 29(5) (2018), 2185–2195.

12.

McCollough

C.H.

, Yu

, Kofler

J.M.

, et al., Degradation of CT low-contrast spatial resolution due to the use of iterative reconstruction and reduced dose levels, Radiology 276(2) (2015), 499–506.

13.

Padole

, Singh

, Lira

, et al., Assessment of filtered back projection, adaptive statistical, and model-based iterative reconstruction for reduced dose abdominal computed tomography, Journal of Computer Aassisted Tomography 39(4) (2015), 462–467.

14.

Sagara

, Hara

A.K.

, Pavlicek

, et al., Abdominal CT: Comparison of low-dose CT with adaptive statistical iterative reconstruction and routine-dose CT with filtered back projection in 53 patients, American Journal of Roentgenology 195(3) (2010), 713–719.

15.

Pooler

B.D.

, Lubner

M.G.

, Kim

D.H.

, et al., Prospective evaluation of reduced dose computed tomography for the detection of low-contrast liver lesions: direct comparison with concurrent standard dose imaging, Eur Radiol 27(5) (2017), 2055–2066.

16.

Goodenberger

M.H.

, Wagner-Bartak

N.A.

, Gupta

, et al., Computed tomography image quality evaluation of a new iterative reconstruction algorithm in the abdomen (Adaptive Statistical Iterative Reconstruction-V) a comparison with model-based iterative reconstruction, adaptive statistical iterative reconstruction, and filtered back projection reconstructions, Journal of Computer Assisted Tomography 42(2) (2018), 184–190.

17.

, Thibault

J.B.

, Bouman

C.A.

, et al., Fast model-based X-ray CT reconstruction using spatially nonhomogeneous ICD optimization, IEEE Transactions on Image Processing 20(1) (2011), 161–175.

18.

Thibault

J.B.

, Sauer

K.D.

, Bouman

C.A.

and Hsieh

, A three-dimensional statistical approach to improved image quality for multislice helical CT, Medical Physics 34(11) (2007), 4526–4544.

19.

Lim

, Kwon

, Cho

, et al., Initial phantom study comparing image quality in computed tomography using adaptive statistical iterative reconstruction and new adaptive statistical iterative reconstruction V, Journal of Computer Assisted Tomography 39(3) (2015), 443–448.

20.

Gatti

, Marchisio

, Fronda

, et al., Adaptive statistical iterative reconstruction–V versus adaptive statistical iterative reconstruction, Journal of Computer Assisted Tomography 42(2) (2018), 191–196.

21.

Jensen

C.T.

, Wagner-Bartak

N.A.

, Vu

L.N.

, et al., Detection of colorectal hepatic metastases is superior at standard radiation dose CT versus reduced dose CT, Radiology 290(2) (2019), 400–409.

22.

, Feng

and Gao

, Deep-learning-based image reconstruction for compressed ultrafast photography, Optics Letters 45(16) (2020), 4400–4403.

23.

Akagi

, Nakamura

, Higaki

, et al., Deep learning reconstruction improves image quality of abdominal ultra-high-resolution CT, Eur Radiol 29(11) (2019), 6163–6171.

24.

Higaki

, Nakamura

, Zhou

, et al., Deep learning reconstruction at CT: phantom study of the image characteristics, Academic Radiology 27(1) (2020), 82–87.

25.

Tatsugami

, Higaki

, Nakamura

, et al., Deep learning-based image restoration algorithm for coronary CT angiography, Eur Radiol 29(10) (2019), 5322–5329.

26.

Singh

, Digumarthy

S.R.

, Muse

V.V.

, et al., Image quality and lesion detection on deep learning reconstruction and iterative reconstruction of submillisievert chest and abdominal CT, American Journal of Roentgenology 214(3) (2020), 566–573.

27.

Kim

J.H.

, Yoon

H.J.

, Lee

, et al., Validation of deep-learning image reconstruction for low-dose chest computed tomography scan: Emphasis on image quality and noise, Korean Journal of Radiology 22(1) (2020), 131–138.

28.

Benz

D.C.

, Benetos

, Rampidis

, et al., Validation of deep-learning image reconstruction for coronary computed tomography angiography: Impact on noise, image quality and diagnostic accuracy, Journal of Cardiovascular Computed Tomography 14(5) (2020), 444–451.

29.

Hong

J.H.

, Park

E.A.

, Lee

, et al., Incremental image noise reduction in coronary CT angiography using a deep learning-based technique with iterative reconstruction, Korean Journal of Radiology 21(10) (2020), 1165–1177.

30.

Narita

, Nakamura

, Higaki

, et al., Deep learning reconstruction of drip-infusion cholangiography acquired with ultra-high-resolution computed tomography, Abdominal Radiology 45(9) (2020), 2698–2704.