Introduction

Ultra-high-resolution computed tomography (U-HRCT), commercially available since 2017, features a smaller detector element and tube focus size than conventional CT. U-HRCT yields images of higher spatial resolution; their usefulness for the examination of the lungs, coronary arteries, and peripheral arteries has been reported [1,2,3,4]. However, compared to conventional CT, U-HRCT has greater image noise due to relatively insufficient incident photons on smaller detectors. As a consequence, increased noise may prevent implementation for abdominal examinations [1, 4, 5].

Model-based iterative reconstruction (MBIR) can improve the image quality and potentially reduce radiation dose [6,7,8]. However, MBIR images are remarkably degraded due to low-frequency noise, particularly at low radiation dose settings [9,10,11,12]. In addition, the MBIR approach usually requires higher computational power and longer computational time. Hybrid iterative reconstruction (hybrid-IR) is faster than MBIR. However, the overall imaging performance of hybrid-IR is not as good as of MBIR in terms of the noise characteristics, spatial resolution, and artifact reduction [6, 13,14,15,16].

Deep learning reconstruction (DLR) (Advanced Intelligent Clear-IQ Engine [AiCE], Canon Medical Systems) is the first commercialized deep learning reconstruction tool. It incorporates deep convolutional neural networks (DCNN) restoration process into the reconstruction flow. For the deep learning-based approach, given hybrid-IR images and high-dose MBIR images as training pairs, statistical features that differentiate signal from the noise and artifacts could be “learned” in the training process and then be “updated” in the DCNN kernel for future inference use (Fig. 1). As DCNN kernel is trained with ideal MBIR images, we expect to see that not only the DLR approach could generate comparable image quality to the MBIR image but also that it takes shorter processing time than MBIR; therefore, DLR could be useful for reconstructing abdominal U-HRCT images. In this study, we investigated the clinical applicability of DLR and compared its image quality to hybrid-IR and MBIR.

Fig. 1
figure 1

Training and inference flowchart of deep learning reconstruction (DLR) algorithm. a Parameters in the DCNN kernel optimized in the training process when differences (loss) between the DCNN output and the ideal MBIR image are stably minimized. b The DLR for the data inference includes three major parts: data domain filtering, hybrid-IR reconstruction, and a DCNN-based restoration module

Full size image

Materials and methods

This retrospective study was approved by our institutional review board; prior informed patient consent was waived because this study used existing CT images including raw data. Patient records and information were anonymized and de-identified prior to analysis.

Study population

We estimated the sample size needed to detect a difference between CT images reconstructed with hybrid-IR, MBIR, and DLR as 20 patients with an effect size of 0.85, an α of 0.013 (one third value of 0.05 because of Bonferroni correction for multiple comparison), and a statistical power of 0.8 [17]. Effect size was calculated based on the preliminary data using images of 10 patients not included in this study.

This study included 46 consecutive patients (31 men, 15 women, age range 34–86 years; mean age, 73 years) who had undergone hepatic dynamic CT using U-HRCT at our institution between December 2017 and April 2018. The clinical indication for hepatic dynamic CT was follow-up after surgery for malignant liver tumor (n = 25), evaluation after chemotherapy for a malignant liver tumor (n = 14), staging of a suspected malignant liver tumor (n = 1), and screening for liver tumors (n = 6).

CT image acquisition

Images were acquired on a U-HRCT scanner (Aquilion Precision, Canon Medical Systems). The scanning protocols were as follows: rotation time 0.75 s, pitch factor 0.806, scanning field of view 40 cm, voltage 120 kV, and tube current 250 mA. Hepatic dynamic CT images were obtained during the hepatic arterial and the equilibrium phase (HAP, EP) in super-high-resolution mode (1792 channels per detector row, 0.25 mm × 160 rows; matrix size, 1024) because these phases are essential for diagnosing hepatocellular carcinoma [18,19,20,21]. An automatic bolus-tracking program was used to time the start of scanning for each phase after contrast medium injection. The trigger threshold level was set at 200 Hounsfield units (HU) in the abdominal aorta at the L1 vertebral body level. HAP and EP scans were started at 17 and 152 s after triggering. The contrast material (600 mgI/kg body weight) was administered using a power injector (Dual Shot, Nemoto Kyorindo) and a 20-gauge catheter inserted into an antecubital vein. The injection duration was 30 s in all patients and the delivery of contrast material was followed by flushing with 30 mL of physiologic saline at the same injection rate.

Although pre-enhanced and portal venous phase (PVP) scans were obtained for the clinical studies, they were not evaluated in ours because they were not performed in super-high-resolution mode.

To assess radiation exposure, we reviewed the CT dose index (CTDIvol) and the dose-length product (DLP) recorded as Digital Imaging and Communications in Medicine (DICOM) data. We also calculated the size-specific dose estimate (SSDE), an index in which the CTDI is corrected by the body habitus [22, 23]. Size-dependent conversion factors were obtained from AAPM Report 204 [24]; they were based on the sum of the anteroposterior and lateral dimensions at the mid-liver level of each patient.

Image analysis

The CT images at HAP and EP were reconstructed with hybrid-IR (Adaptive Iterative Dose Reduction 3-Dimensional [AIDR3D, standard setting]; Canon Medical Systems), MBIR (forward-projected model-based iterative reconstruction solution [FIRST]; Canon Medical Systems), and DLR (AiCE).

Qualitative image analysis

Two board-certified radiologists (Y.N. and K.A. with 14 and 31 years of experience in radiology, respectively) performed consensual qualitative analysis of the CT images. They inspected a total of 276 scans (46 × 2 × 3) of 0.25 mm section thickness that were reconstructed with hybrid-IR, MBIR, and DLR. They were blinded to all patient demographics and CT parameters. The images were presented in random order on a preset soft tissue window; the window width and level were 300 and 60 HU, respectively.

The readers were given standardized instructions and trained on image sets from five patients not included in this study. They ranked the images obtained from the 46 patients for vessel conspicuity (visibility of small structures, especially the depiction of the segmental branch level of the hepatic artery) on HAP images and for overall image quality on HAP and EP images. Vessel grading was on the 5-point Likert scale, where 1 = very poor, 2 = suboptimal, 3 = acceptable, 4 = above average, and 5 = excellent [25]. The overall image quality was also scored on the 5-point Likert scale [26, 27], where 1 = unacceptable diagnostic image quality, 2 = subdiagnostic, 3 = average, 4 = above average, and 5 = excellent [25].

Quantitative image analysis

Quantitative analysis of transverse images (section thickness 0.25 mm) was performed by one radiologist (M.A. with 5 years of experience in radiology). For attenuation measurements, regions of interest (ROIs) were placed within the aorta, portal vein, liver, and paraspinal muscle. Aortic attenuation was recorded at the celiac artery level using a single, manually drawn ROI as large as the vessel lumen; it avoided calcifications and/or soft plaques on the aortic wall. Portal vein attenuation was also recorded based on a single, hand-drawn ROI placed at the right and left portal vein confluence level. Liver attenuation was recorded as the mean measurement value of 4 ROIs in the right anterior, right posterior, left medial, and left lateral segment of the liver. Areas of focal changes in hepatic parenchymal attenuation, large vessels, and prominent artifacts, if any, were carefully avoided. Attenuation of the paraspinal muscle was recorded, also avoiding macroscopic fat infiltration, at the level of the right portal vein. Each value was calculated by averaging the three time measurements. The standard deviation (SD) of attenuation measured in the paraspinal muscle was used as the image noise.

For each of the image sets, the aortic, portal vein, and liver contrast-to-noise ratio (CNR), relative to the muscle, was calculated using the equation:

$$ \mathrm{CNR}=\left({\mathrm{ROI}}_{\mathrm{ORGAN}}-{\mathrm{ROI}}_{\mathrm{MUSCLE}}\right)/N, $$

where ROIORGAN is the mean attenuation of the organ of interest, ROIMUSCLE the mean attenuation of the paraspinal muscle, and N is the noise.

Statistical analysis

Statistically significant differences were evaluated with JMP10 software (SAS Institute). Differences among CT images subjected to hybrid-IR, MBIR, and DLR were determined. The two-sided Wilcoxon signed rank test with Bonferroni correction was applied to examine intergroup differences. Differences of p < 0.013 for multiple comparisons using Bonferroni correction were considered statistically significant.

For qualitative analysis, we calculated interobserver agreement using the weighted kappa statistic to evaluate agreement between the two readers. A kappa statistic in the range of 0.81–1.00 was interpreted as excellent, 0.61–0.80 as substantial, 0.41–0.60 as moderate, 0.21–0.40 as fair, and 0.00–0.20 as poor agreement [28].

Results

Quantitative analysis of the image noise and CNR

As shown in Table 1, image noise was much lower on DLR than hybrid-IR and MBIR images at HAP (median image noise was 24.9, 22.2, and 13.9 HU for hybrid-IR, MBIR, and DLR images, respectively; p < 0.01 for both), while there was a little difference between hybrid-IR and MBIR images (median image noise was 24.9 and 22.2 HU for hybrid-IR and MBIR images, respectively; p < 0.01). On EP images, DLR also yielded significantly lower image noise than hybrid-IR and MBIR (median image noise was 25.5, 23.1, and 14.6 HU for hybrid-IR, MBIR, and DLR images, respectively; p < 0.01 for both); it was lower on MBIR than hybrid-IR images (median image noise was 25.5 and 23.1 HU for hybrid-IR and MBIR images, respectively; p < 0.01). While there was a little difference in attenuation value of each organ among three reconstruction methods both on HAP and EP images (Supplemental Table 1), the aortic, portal vein, and liver CNR was significantly higher with DLR than the other reconstruction methods both on HAP and EP images (p < 0.01). The difference for liver CNR on HAP images and the aortic, portal vein, and liver CNR on EP images between hybrid-IR and MBIR images was not significant.

Table 1 Image noise and CNR on hybrid-IR, MBIR, and DLR images
Full size table

Qualitative analysis

DLR showed the highest overall image scores, and MBIR the lowest scores for both HAP and EP images (Table 2, Figs. 2, 3, and 4). All DLR images had a score of 3 (average) or higher in terms of overall image quality. On the other hand, 17.4% of hybrid-IR and 43.5% of MBIR images were subdiagnostic (score = 2) at HAP. This was true for 34.8% of hybrid-IR and 76.1% of MBIR images at EP. For vessel conspicuity, MBIR images yielded the highest scores; the scores for DLR and hybrid-IR images were comparable (Fig. 5). Interobserver agreement between the two readers was substantial (kappa value range 0.71–0.80).

Table 2 Qualitative analysis scores of hybrid-IR, MBIR, and DLR images
Full size table
Fig. 2
figure 2

Hepatic arterial (a–c) and equilibrium phase images (d–f) of a 72-year-old woman. Reconstruction was with hybrid-IR (a, d), MBIR (b, e), and DLR (c, f). Compared with the hybrid-IR image, the image noise was not reduced on the MBIR image. On the image reconstructed with DLR, the image noise was lower than on the hybrid-IR image

Full size image
Fig. 3
figure 3

Hepatic arterial (a–c) and equilibrium phase images (d–f) of a 76-year-old man. Reconstruction was with hybrid-IR (a, d), MBIR (b, e), and DLR (c, f). The image noise was lower on the DLR image than on the other images

Full size image
Fig. 4
figure 4

Hepatic arterial phase images of a 59-year-old man. Reconstruction was with hybrid-IR (a), MBIR (b), and DLR (c). The image noise on the DLR image was markedly lower than on the other images. Note the improvement of image quality and better delineation not only for the liver but also other organs such as the pancreas and kidney on DLR image compared with hybrid-IR and MBIR images

Full size image
Fig. 5
figure 5

Hepatic arterial phase images of an 81-year-old woman. Axial (a–c) and curved multiplanar reformation (CPR) images (d–f). Reconstruction was with hybrid-IR (a, d), MBIR (b, e), and DLR (c, f). On the hybrid-IR and DLR images, the branch level of the hepatic artery (arrow) was visualized, but the vascular edge was blurred and irregular; it was sharp on the MBIR image

Full size image

Radiation exposure

The median CTDIvol, DLP, and SSDE values for hepatic dynamic scans were 12.6 mGy (range 9.7–16.1), 1240.5 mGy cm (range 910.7–1950.5), and 17.7 mGy (range 15.6–20.4), respectively. They were slightly lower compared to conventional hepatic dynamic CT reported as the Japanese diagnostic reference levels [29].

Discussion

We found that DLR yielded a significantly lower image noise and higher CNRs than hybrid-IR and MBIR at both HAP and EP. The subjective overall image quality score was significantly better with DLR than hybrid-IR and MBIR. Thus, we concluded that DLR can yield better image quality than hybrid-IR and MBIR.

Although a reduction in the slice thickness degrades image quality, we selected 0.25 mm, the thinnest, to maximize the spatial resolution on these scans [30, 31]. In addition, image noise was higher on U-HRCT than conventional HRCT images due to the smaller size of the detectors [1, 4]. Thus, our protocol, thin slice images using U-HRCT, results in worse image quality due to increased noise as compared to conventional images. Indeed, hybrid-IR and MBIR images were graded as subdiagnostic (overall image quality score of 2 or lower) for some cases (between 17.4 and 76.1%). On the other hand, overall image quality score of 3 (average) or higher was assigned for all DLR images for all cases. Based on our findings, we suggest that as DLR allows thinner slices in abdominal U-HRCT images while maintaining image quality, DLR appears as an essential reconstruction method for U-HRCT scanning.

MBIR images yielded significantly lower image noise than hybrid-IR. However, MBIR resulted in the lowest scores among the three reconstruction methods for overall image quality. As determination of the SD is easy and quick, it is widely used to estimate the image noise on CT scans [32, 33]. However, it yields a very limited description of the noise characteristics because two images with very different noise textures may exhibit an identical SD [34]. MBIR is able to reduce many of the high-frequency but not the low-frequency noise components [9,10,11]. Therefore, we supposed that their qualitative image quality was not improved on MBIR images because the low-frequency noise components were not reduced.

In our study, vessel conspicuity was better on MBIR than hybrid-IR and DLR images. MBIR yields better visualization of small vessels than hybrid-IR because it reduces image noise while maintaining image contrast and resolution [16, 25, 35]. The DLR algorithm we applied was developed for abdominal images and not optimized for the evaluation of vessels. Thus, we think that DLR used in this study and MBIR are complementary for U-HRCT of the abdomen. Different DLR algorithms might be needed for the evaluation of vessels.

There was a significant difference in attenuation value of each organ among the three reconstruction methods although the difference was small. Regarding the attenuation value of aorta at HAP, for which the largest difference was observed, MBIR showed the highest value (about 10 HU higher compared to both hybrid-IR and DLR) as already reported [36, 37]. However, DLR yielded the higher CNR compared to the other two reconstruction methods for all organs including the aorta at both phases, indicating that noise reduction with DLR may overcome the difference in attenuation value of each organ.

Although the radiation exposure must be scrutinized [38], dose reduction must be balanced against an acceptable level of diagnostic accuracy. According to Pickhardt et al [12], the ability to detect focal hepatic lesions was reduced on low-dose MBIR images. DCNN for DLR was trained with MBIR images acquired with a sufficient radiation dose. To ascertain the ability to detect focal hepatic lesions on low-dose DLR images, further studies are needed.

Our study has some limitations. The study population was relatively small, and our investigation was retrospective and carried out at a single institution. Therefore, we consider our findings preliminary. We enrolled 46 patients although only 20 were required based on the power calculation. Very large samples may reject null hypotheses with clinically negligible differences, leading that what is insignificant may become significant [39]. However, the difference in image noise between DLR and other algorithms was larger compared to standard deviation of image noise of each reconstruction method, indicating that a significant difference in our data was not clinically negligible even though patient population was slightly larger compared to the required one based on the power calculation. Also, to avoid excessive radiation doses [29], PVP images of our patients were not performed in super-high-resolution mode and, consequently, were not included in this study. However, as PVP imaging is important for the evaluation of hypovascular hepatic metastases and of abnormalities of the portal venous system [40], further studies including PVP images are needed. We did not specifically evaluate focal hepatic lesions, which should be done. Finally, there is no external and further investigation including a larger cohort study and a phantom study is certainly needed.

In conclusion, image noise, overall image quality, and CNR for hepatic U-HRCT images improved with DLR compared to hybrid-IR and MBIR. Thus, DLR can improve abdominal CT image quality using U-HRCT.