Abstract
Purpose
To evaluate the effect of a deep learning–based reconstruction (DLR) method on the conspicuity of hypovascular hepatic metastases on abdominal CT images.
Materials and Methods
This retrospective study with institutional review board approval included 58 patients with hypovascular hepatic metastases. A radiologist recorded the standard deviation of attenuation in the paraspinal muscle as the image noise and the contrast-to-noise ratio (CNR). CNR was calculated as region of interest ([ROI]L − ROIT)/N, where ROIL is the mean liver parenchyma attenuation, ROIT, the mean tumor attenuation, and N, the noise. Two other radiologists graded the conspicuity of the liver lesion on a five-point scale where 1 is unidentifiable and 5 is detected without diagnostic compromise. Only the smallest liver lesion in each patient, classified as smaller or larger than 10 mm, was evaluated. The difference between hybrid iterative reconstruction (IR) and DLR images was determined by using a two-sided Wilcoxon signed-rank test.
Results
The image noise was significantly lower, and the CNR was significantly higher on DLR images than hybrid IR images (median image noise: 19.2 vs 12.8 HU, P < .001; median CNR: tumors < 10 mm: 1.9 vs 2.5; tumors > 10 mm: 1.7 vs 2.2, both P < .001). The scores for liver lesions were significantly higher for DLR images than hybrid IR images (P < .01 for both in tumors smaller or larger than 10 mm).
Conclusion
DLR improved the quality of abdominal CT images for the evaluation of hypovascular hepatic metastases.
© RSNA, 2019
Summary
A deep learning–based reconstruction method can quantitatively and qualitatively improve the quality of abdominal CT images for the evaluation of hypovascular hepatic metastases.
Key Point
■ Compared with hybrid iterative reconstruction (IR) alone, deep learning–based reconstruction after hybrid IR reduced the image noise and improved the quality of abdominal CT images for the evaluation of hypovascular hepatic metastases.
Introduction
Metastatic lesions are more common than primary tumors of the liver (1–3). As surgical resection has decreased the mortality rate and improved the prognosis of patients with hepatic metastases from colorectal cancer, the number of lesions and their size and location are assessed on images (4). CT studies that evaluate the liver, abdomen, and chest simultaneously in a single session are the primary modality for conducting follow-up examinations and for determining the disease stage. However, as the detectability of hepatic metastases, especially those smaller than 10 mm, on contrast material–enhanced CT images remains limited (5–7), further improvements are needed.
Image reconstruction algorithms are applied to reconstruct images with the lowest possible noise without sacrificing their accuracy and spatial resolution. Model-based iterative reconstruction (MBIR), an advanced reconstruction algorithm for CT studies, can improve the image quality and allow for radiation dose reduction (8–10). However, the improved detectability of low-contrast lesions on MBIR images, particularly at low-dose tube flux levels and in larger patients, remains to be demonstrated (11–13). While the MBIR approach tends to require high computational power and longer time, hybrid iterative reconstruction (IR) is faster and more widely used, although its overall imaging performance is inferior to MBIR in terms of noise and artifact reduction (8,14–16). Consequently, next-generation image reconstruction techniques are needed to acquire excellent images at low radiation doses and at acceptable computational costs.
Unlike task-specific algorithms, deep learning belongs to a family of machine learning methods that are based on learning data representations (17,18); deep learning has been used in computer vision tasks and in medical imaging applications (19–23). Deep learning–based reconstruction (DLR) is a method that applies a deep learning–based approach trained on high-quality CT images (24). We expect to see whether DLR can yield high-quality CT images not only at reduced radiation doses but also at shorter processing times than those required for MBIR.
This study evaluated the effect of our DLR method on the conspicuity of hypovascular hepatic metastases on abdominal CT images.
Materials and Methods
The manuscript editing of this study was supported by Canon Medical Systems and Canon Medical Research USA. Some authors are employees of Canon Medical Systems and Canon Medical Research USA; authors not employed by Canon Medical Systems and Canon Medical Research USA controlled the inclusion of data and information that might present a conflict of interest for authors who are employees of Canon Medical Systems and Canon Medical Research USA.
This retrospective study was approved by our institutional review board; prior informed patient consent was waived. Patient records and information were strongly de-identified prior to analysis.
Study Population
A two-sided Wilcoxon signed-rank test was used to estimate the sample size required for the detection of image noise differences on hybrid IR and DLR images. The conditions for sample size calculation were an effect size of 0.5, an α of .05, and a statistical power of 0.95. The effect size was calculated based on preliminary data using images in 10 patients not included in this study.
The inclusion criteria for this retrospective study were the presence of hypovascular hepatic metastases and the availability of hepatic contrast-enhanced CT images obtained at our institution between September 2015 and September 2017. Patients with more than 16 hepatic lesions were excluded because it was burdensome to confirm all lesions. The body mass index (BMI) of each patient was recorded.
One reader (Y.N., with 14 years of experience in abdominal radiology and excluded from qualitative image analyses) confirmed the presence and location of hepatic metastases on contrast-enhanced CT images, PET images obtained by using fluorine 18 fluorodeoxyglucose, and gadoxetic acid–enhanced MR images. Pathologic findings obtained at definitive surgery were also considered.
CT Image Acquisition and Image Reconstruction
CT images were acquired with a 320–detector row scanner (Aquilion One; Canon Medical Systems, Otawara, Japan). The scanning protocol was a rotation time of 0.5 second, beam collimation of 80 × 0.5 mm, reconstruction section thickness and interval of 1.0 mm, pitch factor of 0.813, table movement of 65 mm/sec, default field of view (FOV) of 40 cm, 120 kV, and automatic exposure control (standard deviation of 12). Contrast-enhanced CT images were obtained 100 seconds after the start of contrast material administration (dose, 600 milligrams of iodine per kilogram of body weight). Contrast material (2.0 mL/sec) was injected with a 20-gauge catheter inserted into an antecubital vein by using a power injector (Dual Shot; Nemoto Kyorindo, Tokyo, Japan).
The CT images were reconstructed with DLR and hybrid IR (Adaptive Iterative Dose Reduction with three-dimensional processing, AIDR 3D, standard setting; Canon Medical Systems). For details on DLR, see Appendix E1 (supplement).
Qualitative Image Analysis
Two board-certified radiologists (F.T. and K.A., with 17 and 31 years of experience in radiology, respectively) independently performed a blinded qualitative analysis of the CT images.
In each session, 58 hybrid IR and 58 DLR images were presented in a random order to each reader; they were blinded to patient demographics and CT parameters. The images were initially presented on a preset soft-tissue window (width, 400 HU; level, 40 HU); the readers were allowed to change the display window setting.
Both readers received standardized instructions and were trained on image sets from five patients not included in this study. They rated the overall image quality independently based on a four-point scale (25) where 1 is not evaluable, very high image noise and marked distortion of spatial or contrast resolution: not willing to accept in clinical situations; 2 is poor: willing to accept only in certain clinical situations; 3 is fair: willing to accept in most clinical situations; and 4 is good: willing to accept in all clinical situations (25–27). The readers then assigned a score for contour definition and structure delineation of in-plane blood vessels plus the liver margin on a four-point Likert scale, where 1 is severe blurring, edge definition very poor, margins difficult to discern; 2 is moderate blurring, edge definition poor, margins discernible; 3 is minimal blurring, edge definition good, margins easily discerned; and 4 is no blurring, edges well defined, margins crisp (27).
Hypovascular metastatic liver tumors were classified as smaller or larger than 10 mm. The 10-mm threshold was applied because the detectability of hepatic metastases, especially those smaller than 10 mm, on contrast-enhanced CT remains limited (5–7,28). Both radiologists also recorded the conspicuity of hypovascular metastatic liver tumors using a five-point Likert scale where 1 is definite artifact mimicking a lesion; 2 is probable artifact mimicking a lesion; 3 is subtle lesion; 4 is well-visualized lesion with poorly visualized margins; and 5 is well-visualized lesion with visualized margins (29). As it is more difficult to detect small tumors than large tumors, the visualization of small tumors is of clinical importance and requires improvement. In patients with multiple hepatic metastases, we only evaluated the smallest tumor.
Quantitative Image Analysis
Quantitative analysis was performed by one board-certified abdominal radiologist (Y.N.) on transverse images (section thickness, 1 mm). In patients with multiple hepatic metastases, only the smallest tumor was evaluated. To measure attenuation, a region of interest (ROI) was placed on the tumor, adjacent hepatic parenchyma, and paraspinal muscle. The average ROI size was 143.9 mm2 ± 46.1 (standard deviation [SD]) (range, 55.5–256.2 mm2) for the adjacent hepatic parenchyma and 543.2 mm2 ± 254.7 (range, 189.1– 773.8 mm2) for the paraspinal muscle. Each measurement was repeated three times. The SD of the attenuation measured in the paraspinal muscle was used as the image noise.
The contrast-to-noise ratio (CNR) was calculated as CNR = (ROIL − ROIT)/N, where ROIL is the mean attenuation of the hepatic parenchyma, ROIT, the mean attenuation of the tumor, and N, the noise.
Differences between hybrid IR and DLR images with regard to the image noise and CNR were calculated as the DLR value minus the hybrid IR value.
Radiation Dose
To assess radiation exposure, the volume CT dose index (CTDIvol) and the dose–length product (DLP) displayed on the scanner console were recorded. The size-specific dose estimate (SSDE), an index in which the CTDIvol is corrected by the body habitus, was also calculated (30,31). Size-dependent conversion factors were obtained from the American Association of Physicists in Medicine Report 204 (32); they were based on the sum of the anteroposterior and lateral dimensions at the mid-liver level of each patient.
As not only the radiation dose but also the patient body size is highly correlated with the image quality, we calculated the CTDIvol/BMI value as the indicator of the radiation dose delivered per patient body weight. We determined the effect of the CTDIvol/BMI value on the difference in the image noise and CNR on hybrid IR and DLR images.
Statistical Analysis
Statistical analyses were performed with JMP10 software (SAS Institute, Cary, NC). Differences in the image noise, CNR, and ordinal variables on CT images processed with hybrid IR or DLR were determined by using the two-sided Wilcoxon signed-rank test. The Spearman rank correlation was calculated to assess the correlation between the CTDIvol/BMI value and differences in the image noise and CNR. Differences of P < .05 were considered statistically significant.
For the qualitative analysis, we calculated interobserver agreement using the weighted κ statistic to evaluate agreement between the two readers (33).
Results
Patient Population
The sample size needed to detect differences in the image noise between hybrid IR and DLR images was calculated to be 57 patients. On the basis of our criteria, 62 patients were eligible for inclusion in the study; we excluded four of them because they harbored more than 16 metastatic liver lesions. Thus, the final study population consisted of 58 patients (41 male subjects and 17 female subjects; age range, 29–84 years; median age, 68.0 years) (Fig 1). In 12 patients who had undergone partial hepatectomy, the diagnosis of hepatic metastasis was based on pathologic proof of the tumor burden; in the other 46 patients, it was based on tumor growth observed during follow-up. The 58 patients were confirmed to harbor 209 hepatic metastases (size range, 4.1–61.8 mm; median, 12.5 mm; 56 tumors < 10 mm; 153 tumors ≥ 10 mm): 22 patients had solitary lesions and 36 harbored two to 15 lesions (Table 1).
Figure 1:
Flowchart of patient enrollment.
Table 1:
Patient Characteristics
Overall Image Quality
Both readers assigned significantly higher scores to DLR images than hybrid IR images for the overall image quality and for the contour definition and structure delineation of in-plane blood vessels and liver margin (both, P < .001) (Figs 2, 3; Table 2). The median image noise on CT images reconstructed with hybrid IR and DLR was 19.2 and 12.8 HU, respectively. The noise was significantly lower on DLR images than hybrid IR images (P < .001) (Fig 4, Table 3). The median difference in the image noise was −5.3 (95% confidence interval [CI]: −5.9, −4.7). Interobserver agreement was good (κ value range, 0.75–0.91).
Figure 2a:
A 59-year-old man with hepatic metastases from rectal cancer. (a) Hybrid iterative reconstruction (IR) image. (b) Deep learning–based reconstruction (DLR) image. The window settings are the same (window level and window width = 40 and 400 HU, respectively). The arrow points to the hepatic metastatic lesion. The image noise is lower, and the edge and margin of the blood vessel and of the lesion are more clearly visualized on the DLR image than the hybrid IR image.
Figure 3a:
A 50-year-old man with hepatic metastases from rectal cancer. (a) Hybrid iterative reconstruction (IR) image. (b) Deep learning–based reconstruction (DLR) image. The window settings are the same (window level and window width = 40 and 400 HU, respectively). The arrows point to the hepatic metastatic lesions. The image noise is lower, and the lesions are more clearly visualized on the DLR image than the hybrid IR image.
Table 2:
Subjective Overall Image Quality Scores
Figure 4a:
(a) Plot shows image noise on CT images reconstructed with hybrid iterative reconstruction (IR) or deep learning–based reconstruction (DLR). The solid line indicates the median image noise. It is significantly lower on the DLR image than the hybrid IR images (P < .001). (b) Plot shows difference in the image noise on hybrid IR and DLR images (median, −5.3; 95% confidence interval: −5.9, −4.7).
Table 3:
Image Noise and CNR
Figure 2b:
A 59-year-old man with hepatic metastases from rectal cancer. (a) Hybrid iterative reconstruction (IR) image. (b) Deep learning–based reconstruction (DLR) image. The window settings are the same (window level and window width = 40 and 400 HU, respectively). The arrow points to the hepatic metastatic lesion. The image noise is lower, and the edge and margin of the blood vessel and of the lesion are more clearly visualized on the DLR image than the hybrid IR image.
Figure 3b:
A 50-year-old man with hepatic metastases from rectal cancer. (a) Hybrid iterative reconstruction (IR) image. (b) Deep learning–based reconstruction (DLR) image. The window settings are the same (window level and window width = 40 and 400 HU, respectively). The arrows point to the hepatic metastatic lesions. The image noise is lower, and the lesions are more clearly visualized on the DLR image than the hybrid IR image.
Figure 4b:
(a) Plot shows image noise on CT images reconstructed with hybrid iterative reconstruction (IR) or deep learning–based reconstruction (DLR). The solid line indicates the median image noise. It is significantly lower on the DLR image than the hybrid IR images (P < .001). (b) Plot shows difference in the image noise on hybrid IR and DLR images (median, −5.3; 95% confidence interval: −5.9, −4.7).
Lesion Analysis
Only the smallest tumor in each patient was evaluated, and a 10-mm size threshold was applied. Consequently, of the 209 hepatic lesions, we analyzed 58 (range, 4.1–28.9 mm; median, 9.9 mm). Of these, three measured < 5 mm, 28 were ≥ 5 mm and ≤ 10 mm, 23 were ≥ 10 mm and ≤ 20 mm, and four measured ≥ 20 mm. The conspicuity score assigned by both readers was significantly higher for DLR images than hybrid IR images (P < .001 for both in tumors smaller than 10 mm; P < .001 and .001 readers 1 and 2, respectively, in tumors larger than 10 mm; Tables 4, 5). Interobserver agreement was good (κ value range, 0.83–0.93).
Table 4:
Subjective Image Quality Scores for Hypovascular Hepatic Metastatic Lesions Smaller Than 10 mm
Table 5:
Subjective Image Quality Scores for Hypovascular Hepatic Metastatic Lesions Larger Than 10 mm
The CNR for tumors smaller than 10 mm was significantly higher on DLR images than hybrid IR images (median CNR, 2.5 vs 1.9, P < .001; median difference, 0.6 [95% CI: 0.5, 0.8]; Fig 5, Table 3). The CNR for tumors larger than 10 mm also was significantly higher on DLR images than hybrid IR images (median CNR, 2.2 vs 1.7; P < .001; Fig 6, Table 3). The median difference in the CNR was 0.5 (95% CI: 0.4, 0.8).
Figure 5a:
(a) Plot shows the contrast-to-noise ratio (CNR) of tumors smaller than 10 mm on hybrid iterative reconstruction (IR) and deep learning–based reconstruction (DLR) images. The solid line indicates the median CNR. It is significantly higher on DLR images than hybrid IR images (P < .001). (b) Plot shows the difference in the CNR on hybrid IR and DLR images of tumors smaller than 10 mm (median, 0.6; 95% confidence interval: 0.5, 0.8).
Figure 6a:
(a) Plot shows the contrast-to-noise ratio (CNR) of tumors larger than 10 mm on hybrid iterative reconstruction (IR) and deep learning–based reconstruction (DLR) images. The solid line indicates the median CNR. It is higher on DLR images than hybrid IR images (P < .001). (b) Plot shows the difference in the CNR on hybrid IR and DLR images of tumors larger than 10 mm (median, 0.5; 95% confidence interval: 0.4, 0.8).
Figure 5b:
(a) Plot shows the contrast-to-noise ratio (CNR) of tumors smaller than 10 mm on hybrid iterative reconstruction (IR) and deep learning–based reconstruction (DLR) images. The solid line indicates the median CNR. It is significantly higher on DLR images than hybrid IR images (P < .001). (b) Plot shows the difference in the CNR on hybrid IR and DLR images of tumors smaller than 10 mm (median, 0.6; 95% confidence interval: 0.5, 0.8).
Figure 6b:
(a) Plot shows the contrast-to-noise ratio (CNR) of tumors larger than 10 mm on hybrid iterative reconstruction (IR) and deep learning–based reconstruction (DLR) images. The solid line indicates the median CNR. It is higher on DLR images than hybrid IR images (P < .001). (b) Plot shows the difference in the CNR on hybrid IR and DLR images of tumors larger than 10 mm (median, 0.5; 95% confidence interval: 0.4, 0.8).
Effect of Radiation Dose and Patient Body Size
The median CTDIvol, DLP, and SSDE for CT images were 10.3 mGy (range, 5.0–23.8 mGy), 244.4 mGy · cm (range, 121.0–826.0 mGy · cm), and 15.1 mGy (range, 8.4–27.9 mGy), respectively. The median BMI was 22.2 kg/m2 (range, 14.7–33.5 kg/m2).
The median CTDIvol/BMI value was 0.5 (range, 0.3–0.8). There was no significant correlation between the CTDIvol/BMI value and the difference in the image noise. However, there was a significant inverse correlation between the CTDIvol/BMI value and the difference in the CNR (ρ = 0.13 and −0.33; P = .33 and .01 for the difference in the noise and the CNR, respectively; see Fig E8 [supplement]).
Calculation Times
As DLR images were reconstructed on an off-line workstation and hybrid IR and MBIR images were reconstructed on the instrument console, the hardware was different. Therefore, comparing the processing time was difficult. Nonetheless, from the viewpoint of algorithm calculation time, DLR was approximately three to five times faster than MBIR.
Discussion
Difficulties in identifying hypovascular hepatic metastases on contrast-enhanced CT images increase as the image noise increases. Therefore, to improve the detectability of hepatic metastases, the image noise must be reduced (34,35). As we found that the image noise of the paraspinal muscle was significantly lower and the subjective overall image quality was significantly higher on DLR images than hybrid IR images, we hypothesized that the noise on hepatic CT images subjected to DLR would be lower than that on hybrid IR images, resulting in better image quality.
It has been reported (6,7) that the sensitivity of contrast-enhanced CT for depicting hepatic metastases was low, especially for lesions smaller than 10 mm. We found that the conspicuity score and the CNR were significantly higher on DLR images than hybrid IR images of hypovascular metastatic liver tumors smaller and larger than 10 mm. This suggests that DLR is superior to hybrid IR, especially for the assessment of small tumors on CT images.
Although Higaki et al (36) reported that MBIR yielded better image resolution than hybrid IR, the improved detectability of low-contrast lesions on MBIR images remains to be confirmed (13). We found that with regard to the overall image quality and the conspicuity and CNR of hypovascular metastatic hepatic lesions, DLR returned better results than conventional hybrid IR. Our training target dataset for DLR comprised high-quality MBIR images acquired at sufficient radiation doses. Therefore, DLR could generate high-quality images compared with hybrid IR, and this is the first demonstration of using the deep learning approach in learning the performance of MBIR and of its subsequent application to the hybrid IR process.
DLR has other advantages over MBIR and hybrid IR. As the improved detectability of low-contrast lesions acquired at low-dose tube flux levels and in larger patients on commercially available MBIR approaches remains to be demonstrated (11–13), DLR trained on MBIR images may not improve the detectability of low-contrast lesions. However, the MBIR iterations applied for the target at DLR exceeded those of the commercially available MBIR. The datasets used for training the neural network were acquired with a target SD index (5.0–7.0) lower than the standard SD for abdominal CT (37). To obtain images with an SD index of 5.0–7.0, images of patients who were thin were collected, and the delivered radiation dose was high relative to their physique. Taken together, DLR has the potential to achieve better image quality compared with the commercially available MBIR. In addition, the calculation time required for DLR was much shorter than that of MBIR, and the noise reduction with DLR was greater than that of MBIR or hybrid IR. Therefore, DLR may be useful, especially for ultra-high-resolution CT that requires higher computation power (38). In addition, the image noise is substantially increased at low radiation doses, especially when large objects are scanned (39). We found no significant correlation between the CTDIvol/BMI value and the difference in the noise, indicating that DLR yielded similar noise reduction effects irrespective of the radiation dose and the patient habitus. Therefore, we suggest that DLR is a robust reconstruction approach that can yield a consistent image quality for the diagnosis of pediatric and adult patients. Finally, the radiation dose should be as low as possible without degrading the image quality (40). We found a significant inverse correlation between the CTDIvol/BMI value and the difference in the CNR, indicating that an increase in the CNR can be obtained on images acquired at lower radiation doses. Consequently, DLR may be useful for the detection of hepatic lesions on images obtained at lower radiation dose settings.
Our study had some limitations and we consider it to be preliminary. The study population was relatively small, although the power analysis indicated that its size was sufficient for our objective. This was a single-institution retrospective study, and only two readers evaluated the images of metastatic hepatic lesions. As the detectability of hepatic tumors was not evaluated, clinical studies are underway. We did not directly compare the image quality of DLR and MBIR images because MBIR is not routinely used at our institution. Finally, as we only evaluated hypovascular hepatic metastases, additional studies are needed to verify the utility of DLR for the diagnosis of other hepatic tumors such as hepatocellular carcinoma.
In conclusion, the overall image quality, conspicuity, and CNR of hypovascular hepatic lesions on DLR images were superior to those on hybrid IR images even for lesions smaller than 10 mm. DLR resulted in quantitative and qualitative improvements of abdominal CT images acquired for the evaluation of hypovascular hepatic metastases. Future studies will focus on the application of DLR to different clinical scenarios, to different patient populations with a diverse body habitus, and to different acquisition techniques involving low radiation doses and high spatial resolution techniques.
APPENDIX
SUPPLEMENTAL FIGURES
Acknowledgments
Acknowledgments
The authors thank So Tsushima, MS, Takuya Nemoto, MS, and Hiroki Taguchi, MS, for their valuable contributions to this study.
K.A. received a research grant from Canon Medical Systems.
Disclosures of Conflicts of Interest: Y.N. disclosed no relevant relationships. T.H. disclosed no relevant relationships. F.T. disclosed no relevant relationships. J.Z. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: employee of Canon Medical Systems. Other relationships: disclosed no relevant relationships. Z.Y. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: employee of Canon Medical Systems; has patent filed as employee of Canon Medical Research USA. Other relationships: disclosed no relevant relationships. N.A. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: employee of Canon Medical Systems. Other relationships: disclosed no relevant relationships. Y.I. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: employee of Canon Medical Systems. Other relationships: disclosed no relevant relationships. M.I. disclosed no relevant relationships. K.A. Activities related to the present article: institution received research grant from Canon Medical Systems. Activities not related to the present article: institution receives research grants from Hitachi, Fujitsu, Bayer-Yakuhin, Eizai, Daiichi-Sankyo, and Nemoto-Kyorindo; author is paid for lectures by Hitachi, Bayer-Yakuhin, Eizai, and Daiichi-Sankyo. Other relationships: disclosed no relevant relationships.
Abbreviations:
- BMI
- body mass index
- CI
- confidence interval
- CNR
- contrast-to-noise ratio
- CTDIvol
- volume CT dose index
- DLP
- dose–length product
- DLR
- deep learning–based reconstruction
- FOV
- field of view
- IR
- iterative reconstruction
- MBIR
- model-based iterative reconstruction
- ROI
- region of interest
- SD
- standard deviation
- SSDE
- size-specific dose estimate
References
- 1.Ananthakrishnan A, Gogineni V, Saeian K. Epidemiology of primary and secondary liver cancers. Semin Intervent Radiol 2006;23(1):47–63. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Bosch FX, Ribes J, Díaz M, Cléries R. Primary liver cancer: worldwide incidence and trends. Gastroenterology 2004;127(5 Suppl 1):S5–S16. [DOI] [PubMed] [Google Scholar]
- 3.de Ridder J, de Wilt JH, Simmer F, Overbeek L, Lemmens V, Nagtegaal I. Incidence and origin of histologically confirmed liver metastases: an explorative case-study of 23,154 patients. Oncotarget 2016;7(34):55368–55376. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Maher B, Ryan E, Little M, Boardman P, Stedman B. The management of colorectal liver metastases. Clin Radiol 2017;72(8):617–625. [DOI] [PubMed] [Google Scholar]
- 5.Tsurusaki M, Sofue K, Murakami T. Current evidence for the diagnostic value of gadoxetic acid-enhanced magnetic resonance imaging for liver metastasis. Hepatol Res 2016;46(9):853–861. [DOI] [PubMed] [Google Scholar]
- 6.Muhi A, Ichikawa T, Motosugi U, et al. Diagnosis of colorectal hepatic metastases: comparison of contrast-enhanced CT, contrast-enhanced US, superparamagnetic iron oxide-enhanced MRI, and gadoxetic acid-enhanced MRI. J Magn Reson Imaging 2011;34(2):326–335. [DOI] [PubMed] [Google Scholar]
- 7.Niekel MC, Bipat S, Stoker J. Diagnostic imaging of colorectal liver metastases with CT, MR imaging, FDG PET, and/or FDG PET/CT: a meta-analysis of prospective studies including patients who have not previously undergone treatment. Radiology 2010;257(3):674–684. [DOI] [PubMed] [Google Scholar]
- 8.Volders D, Bols A, Haspeslagh M, Coenegrachts K. Model-based iterative reconstruction and adaptive statistical iterative reconstruction techniques in abdominal CT: comparison of image quality in the detection of colorectal liver metastases. Radiology 2013;269(2):469–474. [DOI] [PubMed] [Google Scholar]
- 9.Chang W, Lee JM, Lee K, et al. Assessment of a model-based, iterative reconstruction algorithm (MBIR) regarding image quality and dose reduction in liver computed tomography. Invest Radiol 2013;48(8):598–606. [DOI] [PubMed] [Google Scholar]
- 10.Fontarensky M, Alfidja A, Perignon R, et al. Reduced radiation dose with model-based iterative reconstruction versus standard dose with adaptive statistical iterative reconstruction in abdominal CT for diagnosis of acute renal colic. Radiology 2015;276(1):156–166. [DOI] [PubMed] [Google Scholar]
- 11.Nishizawa M, Tanaka H, Watanabe Y, Kunitomi Y, Tsukabe A, Tomiyama N. Model-based iterative reconstruction for detection of subtle hypoattenuation in early cerebral infarction: a phantom study. Jpn J Radiol 2015;33(1):26–32. [DOI] [PubMed] [Google Scholar]
- 12.Euler A, Stieltjes B, Szucs-Farkas Z, et al. Impact of model-based iterative reconstruction on low-contrast lesion detection and image quality in abdominal CT: a 12-reader-based comparative phantom study with filtered back projection at different tube voltages. Eur Radiol 2017;27(12):5252–5259. [Published correction appears in Eur Radiol 2017;27(12):5260.] 10.1007/s00330-017-4825-9. [DOI] [PubMed] [Google Scholar]
- 13.Racine D, Ba AH, Ott JG, Bochud FO, Verdun FR. Objective assessment of low contrast detectability in computed tomography with Channelized Hotelling Observer. Phys Med 2016;32(1):76–83. [DOI] [PubMed] [Google Scholar]
- 14.Yasaka K, Furuta T, Kubo T, et al. Full and hybrid iterative reconstruction to reduce artifacts in abdominal CT for patients scanned without arm elevation. Acta Radiol 2017;58(9):1085–1093. [DOI] [PubMed] [Google Scholar]
- 15.Nakamoto A, Kim T, Hori M, et al. Clinical evaluation of image quality and radiation dose reduction in upper abdominal computed tomography using model-based iterative reconstruction; comparison with filtered back projection and adaptive statistical iterative reconstruction. Eur J Radiol 2015;84(9):1715–1723. [DOI] [PubMed] [Google Scholar]
- 16.Deák Z, Grimm JM, Treitl M, et al. Filtered back projection, adaptive statistical iterative reconstruction, and a model-based iterative reconstruction in abdominal CT: an experimental clinical study. Radiology 2013;266(1):197–206. [DOI] [PubMed] [Google Scholar]
- 17.Bengio Y, Courville A, Vincent P. Representation learning: a review and new perspectives. IEEE Trans Pattern Anal Mach Intell 2013;35(8):1798–1828. [DOI] [PubMed] [Google Scholar]
- 18.LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521(7553):436–444. [DOI] [PubMed] [Google Scholar]
- 19.Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556. [preprint] https://arxiv.org/abs/1409.1556. Posted 2015. Accessed May 30, 2018.
- 20.He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. arXiv:1512.03385. [preprint] https://arxiv.org/abs/1512.03385. Posted 2015. Accessed May 30, 2018.
- 21.Powell S, Magnotta VA, Johnson H, Jammalamadaka VK, Pierson R, Andreasen NC. Registration and machine learning-based automated segmentation of subcortical and cerebellar brain structures. Neuroimage 2008;39(1):238–247. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Roth HR, Lu L, Seff A, et al. A new 2.5D representation for lymph node detection using random sets of deep convolutional neural network observations. Med Image Comput Assist Interv 2014;17(Pt 1):520–527. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Qi Dou, Hao Chen, Lequan Yu, et al. Automatic detection of cerebral microbleeds from MR images via 3D convolutional neural networks. IEEE Trans Med Imaging 2016;35(5):1182–1195. [DOI] [PubMed] [Google Scholar]
- 24.Zhu B, Liu JZ, Cauley SF, Rosen BR, Rosen MS. Image reconstruction by domain-transform manifold learning. Nature 2018;555(7697):487–492. [DOI] [PubMed] [Google Scholar]
- 25.Likert R. A technique for the measurement of attitudes. Arch Psychol 1932;22(140):55. https://psycnet.apa.org/record/1933-01885-001. Accessed May 30, 2018. [Google Scholar]
- 26.Phelps AS, Naeger DM, Courtier JL, et al. Pairwise comparison versus Likert scale for biomedical image assessment. AJR Am J Roentgenol 2015;204(1):8–14. [DOI] [PubMed] [Google Scholar]
- 27.Shuman WP, Chan KT, Busey JM, et al. Standard and reduced radiation dose liver CT images: adaptive statistical iterative reconstruction versus model-based iterative reconstruction-comparison of findings and image quality. Radiology 2014;273(3):793–800. [DOI] [PubMed] [Google Scholar]
- 28.Mainenti PP, Mancini M, Mainolfi C, et al. Detection of colo-rectal liver metastases: prospective comparison of contrast enhanced US, multidetector CT, PET/CT, and 1.5 Tesla MR with extracellular and reticulo-endothelial cell specific contrast agents. Abdom Imaging 2010;35(5):511–521. [DOI] [PubMed] [Google Scholar]
- 29.Lee KH, Lee JM, Moon SK, et al. Attenuation-based automatic tube voltage selection and tube current modulation for dose reduction at contrast-enhanced liver CT. Radiology 2012;265(2):437–447. [DOI] [PubMed] [Google Scholar]
- 30.Brady SL, Kaufman RA. Investigation of American Association of Physicists in Medicine Report 204 size-specific dose estimates for pediatric CT implementation. Radiology 2012;265(3):832–840. [DOI] [PubMed] [Google Scholar]
- 31.Christner JA, Braun NN, Jacobsen MC, Carter RE, Kofler JM, McCollough CH. Size-specific dose estimates for adult patients at CT of the torso. Radiology 2012;265(3):841–847. [DOI] [PubMed] [Google Scholar]
- 32.American Association of Physicists in Medicine . Size-specific dose estimates (SSDE) in pediatric and adult body CT examinations (Task Group 204). College Park, MD: American Association of Physicists in Medicine, 2011. https://www.aapm.org/pubs/reports/RPT_204.pdf. Accessed May 30, 2018. [Google Scholar]
- 33.Svanholm H, Starklint H, Gundersen HJ, Fabricius J, Barlebo H, Olsen S. Reproducibility of histomorphologic diagnoses with special reference to the kappa statistic. APMIS 1989;97(8):689–698. [DOI] [PubMed] [Google Scholar]
- 34.Nagayama Y, Iyama A, Oda S, et al. Dual-layer dual-energy computed tomography for the assessment of hypovascular hepatic metastases: impact of closing k-edge on image quality and lesion detectability. Eur Radiol 2019;29(6):2837–2847. [DOI] [PubMed] [Google Scholar]
- 35.Kanal KM, Chung JH, Wang J, et al. Image noise and liver lesion detection with MDCT: a phantom study. AJR Am J Roentgenol 2011;197(2):437–441. [DOI] [PubMed] [Google Scholar]
- 36.Higaki T, Tatsugami F, Fujioka C, et al. Visualization of simulated small vessels on computed tomography using a model-based iterative reconstruction technique. Data Brief 2017;13:437–443. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Kalra MK, Maher MM, Kamath RS, et al. Sixteen-detector row CT of abdomen and pelvis: study for optimization of Z-axis modulation technique performed in 153 patients. Radiology 2004;233(1):241–249. [DOI] [PubMed] [Google Scholar]
- 38.Kakinuma R, Moriyama N, Muramatsu Y, et al. Ultra-high-resolution computed tomography of the lung: image quality of a prototype scanner. PLoS One 2015;10(9):e0137165 [Published correction appears in PLoS One 2015;10(12):e0145357.] 10.1371/journal.pone.0137165. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Qurashi AA, Rainford LA, Alshamrani KM, Foley SJ. The impact of obesity on abdominal CT radiation dose and image quality. Radiat Prot Dosimetry 2018 Dec 1 [Epub ahead of print]. [DOI] [PubMed] [Google Scholar]
- 40.Sagara Y, Hara AK, Pavlicek W, Silva AC, Paden RG, Wu Q. Abdominal CT: comparison of low-dose CT with adaptive statistical iterative reconstruction and routine-dose CT with filtered back projection in 53 patients. AJR Am J Roentgenol 2010;195(3):713–719. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.