The reliability of virtual non-contrast reconstructions of photon-counting detector CT scans in assessing abdominal organs

Abstract

Background

Spectral imaging of photon-counting detector CT (PCD-CT) scanners allows for generating virtual non-contrast (VNC) reconstruction. By analyzing 12 abdominal organs, we aimed to test the reliability of VNC reconstructions in preserving HU values compared to real unenhanced CT images.

Methods

Our study included 34 patients with pancreatic cystic neoplasm (PCN). The VNC reconstructions were generated from unenhanced, arterial, portal, and venous phase PCD-CT scans using the Liver-VNC algorithm. The observed 11 abdominal organs were segmented by the TotalSegmentator algorithm, the PCNs were segmented manually. Average densities were extracted from unenhanced scans (HU_unenhanced), postcontrast (HU_postcontrast) scans, and VNC reconstructions (HU_VNC). The error was calculated as HU_error=HU_VNC–HU_unenhanced. Pearson’s or Spearman’s correlation was used to assess the association. Reproducibility was evaluated by intraclass correlation coefficients (ICC).

Results

Significant differences between HU_unenhanced and HU_{VNC[unenhanced]} were found in vertebrae, paraspinal muscles, liver, and spleen. HU_{VNC[unenhanced]} showed a strong correlation with HU_unenhanced in all organs except spleen (r = 0.45) and kidneys (r = 0.78 and 0.73). In all postcontrast phases, the HU_VNC had strong correlations with HU_unenhanced in all organs except the spleen and kidneys. The HU_error had significant correlations with HU_unenhanced in the muscles and vertebrae; and with HU_postcontrast in the spleen, vertebrae, and paraspinal muscles in all postcontrast phases. All organs had at least one postcontrast VNC reconstruction that showed good-to-excellent agreement with HU_unenhanced during ICC analysis except the vertebrae (ICC: 0.17), paraspinal muscles (ICC: 0.64–0.79), spleen (ICC: 0.21–0.47), and kidneys (ICC: 0.10–0.31).

Conclusions

VNC reconstructions are reliable in at least one postcontrast phase for most organs, but further improvement is needed before VNC can be utilized to examine the spleen, kidneys, and vertebrae.

Background

The technology of photon counting by single photon detectors has been around for decades in quantum information science. However, in the field of human medicine, the first photon-counting detector CT scanner (PCD-CT) has only recently been introduced. This detector system enables acquiring spectral information from the imaged structures that makes the subtraction of iodine from the postcontrast scans possible generating virtual non-contrast (VNC) reconstructions during post-processing [1]. A more detailed description of this technology can be found in the article by Flohr et al. [2].

The introduction of PCD-CT scanners in clinical practice paved the way for clinical research [3]. In abdominal imaging, the assessment of certain findings such as hepatic steatosis [4], chronic calcifying pancreatitis [5], and kidney stones [6] requires unenhanced CT scans before contrast material administration. A common assumption is that reliable VNC reconstructions generated from the postcontrast CT scans could be used in the future to replace real unenhanced phase CT scans reducing the radiation dose of patients.

The reliability of VNC reconstructions of dual-energy CT scanners is a widely researched area [7,8,9], however, only a handful of studies investigated the applicability and reliability of VNC maps of the novel PCD-CT in abdominal imaging involving the analysis of multiple organs [10,11,12]. By the subjective 5-point scale analysis of the image quality, subjective image noise, and noise texture, Mergen et al. reported that VNC reconstructions may have a lower image quality compared to real unenhanced abdominal scans, however with only a moderate interobserver agreement [10]. Previous publications that have quantitatively validated the reliability of VNC reconstructions of PCD-CT scanners have mainly focused on comparing the average Hounsfield units (HU) density values of the investigated organs. Niehoff et al. [11] assessed selected regions of interest (ROIs) of the liver, spleen, kidney, vertebrae, paraspinal muscles, aorta, and fat and reported significant differences between VNC reconstructions vs. real unenhanced scans in both the arterial and venous phase scans for all the investigated organs. In a similar study, Schoenbeck et al. [12] investigated selected ROIs of the liver, spleen, renal cortex, aorta, paraspinal muscles, and subcutaneous fat and they compared the performance of the prior VNC algorithm and the later Liver-VNC algorithm. The authors found significant differences between the real unenhanced scans and their VNC reconstructions in the liver, renal cortex, aorta, paraspinal muscle, and subcutaneous fat, but reported no significant differences in the spleen using the VNC algorithm. Moreover, when comparing the VNC reconstructions of the portal venous phase scans to the real unenhanced ones, the authors found significant differences in all organs except the paraspinal muscles whether the earlier VNC or the later Liver-VNC algorithm was used. However, these previous studies accessed only manually selected circular ROIs of the investigated organs, cystic structures were not evaluated, and it was not assessed whether the error of the reconstruction algorithm shows correlations with the baseline density and the contrast-enhancement of the organs.

In our study, we have enrolled patients diagnosed with pancreatic cystic neoplasm (PCN) as this patient population requires regular follow-up imaging and therefore could benefit significantly from radiation dose reduction if unenhanced phase scans could be replaced with VNC reconstructions in the future. To the best of our knowledge, the reliability of VNC reconstructions of PCD-CT scanners has not yet been investigated in this special population. Investigating PCNs besides other abdominal organs also allows for the reliability assessment of the algorithm in the case of cystic structures and small calcifications. Therefore our study had a special focus on the evaluation of VNC reconstructions’ image quality affecting the radiological assessment of these small calcifications.

We aimed to investigate the reliability of VNC of triple-phasic scans compared to real unenhanced scans. Our study covers the semiquantitative and quantitative evaluation of 12 abdominal organs. We also aimed to investigate the correlation of the density error with the baseline density and the contrast enhancement of organs. Finally, we also aimed to illustrate the most common pitfalls of the reconstruction algorithm, which should be kept in mind when using it.

Methods

Patient population

Our study was approved by the institutional ethics committee. The study adheres to the World Medical Association guidelines and the Declaration of Helsinki, revised in 2000 in Edinburgh. As this is a retrospective study, the need for written informed patient consent was waived by the ethics committee. All patient data were analyzed anonymously.

Follow-up PCD-CT scans from consecutive PCN patients were retrospectively collected between March 2022 – May 2023. Consecutive patients in this time frame were retrospectively identified, who were diagnosed and followed for PCN in the Surgery Department of our University in accordance with the European evidence-based Guidelines of the European Study Group on Cystic Tumours of the Pancreas [13] and who were referred for follow-up CT examination to our Medical Imaging Centre after the installation of the new PCD-CT scanner. CT scans of these patients that were carried out in the considered time frame due to indications other than the regular follow-up were not included. 73 patients were scanned during their follow-up, which led to 86 CTs, whereas 11 patients had two and 1 patient had 3 CT examinations. One scan was excluded due to the incomplete coverage of the PCN. A further six cases were excluded because the PCN could not be contoured due to undefined, blurred tumor margins. The unenhanced phase scans of 44 scans could not be reconstructed with spectral information due to technical issues that resulted in data loss. Therefore, the final patient cohort consisted of 34 patients with 35 CTs. Figure 1 demonstrates the patient selection and data analysis strategy.

Fig. 1

Flowchart on the main steps of the study. PCD-CT: photon-counting detector CT; PCN: pancreatic cystic neoplasm

Imaging protocol

The patients were examined at our Institution with a PCD-CT scanner (NAEOTOM Alpha, VA50; Siemens Healthineers) according to our routine pancreas imaging protocol that included an unenhanced scan, followed by an arterial phase, a portal venous phase, and a venous phase postcontrast scan. The scans were performed with a tube voltage of 120 kVp and an automated tube current modulation. The rotation time was 0.5 s, the pitch was 0.80, the single collimation width was 0.40 mm, the total collimation width was 144 × 0.40 mm, and the reconstruction matrix was 512 × 512 for all scans.

For the contrast-enhanced scans, an iodinated non-ionic contrast agent (either Ultravist 370 or Iomeron 350) was applied using a power injector followed by a 40 mL saline chaser. A power injector that automatically adjusts the injection flow rates by monitoring the peak pressures was used, therefore, the actual contrast injection flow rate was dependent on patient-specific factors such as cardiovascular state, hydration status, etc. resulting in injection flow rates of 2.7–4.4 mL/s in the patient cohort. The amount of contrast agent was adjusted to the patient’s body weight. The amount of iodine injected was 340 ± 52 mg iodine/kg in the patient cohort. Automated timing of the contrast injection was performed with a bolus tracking technique triggered by the peak contrast enhancement measured in the thoracic aorta with a manually placed circular ROI. The timing for the arterial, portal, and venous phase scans were 23 s, 45 s, and 75 s, respectively.

The arterial phase scans were acquired from above the diaphragm to the iliac crest, the portal phase scans were focused on the pancreas covering the upper abdomen, while the unenhanced and the venous phase postcontrast scans were acquired from above the diaphragm to below the symphysis. The field-of-view was manually set by the radiographer to cover the entire body on the axial view.

The CT scans were reconstructed in the axial plane with the so-called spectral post-processing (SPP) reconstruction algorithm of the vendor that preserves all the spectral information. A Qr40 soft kernel with a quantum iterative reconstruction algorithm at a strength level of 3 was used. All scans were reconstructed with a slice thickness of 2.0 mm and an increment of 1.5 mm.

Image postprocessing and segmentation

The VNC images were reconstructed with the Liver-VNC algorithm of the dedicated eXamine research software (Siemens Healthineers, Forchheim, Germany) using the manufacturer’s default settings: the thresholds on high energy bin (65 keV) reconstruction were − 92 HU for fat and 58 HU for soft tissues, these values on low energy bin (20 keV) reconstruction were − 100 HU for fat and 59 HU for soft tissue, while the applied iodine enhancement ratio was 2.0. We used the Liver-VNC algorithm instead of the previous VNC algorithm of the manufacturer. While the previous VNC algorithm performs material decomposition based on the separation of air, iodine, and water; the Liver-VNC algorithm has been developed to take into account fat and soft tissue in addition to the iodine contrast agent to provide a more appropriate algorithm for abdominal organs.

Further processing was completed using the 3D Slicer software v5.2.2 [14]. The covered body regions on the unenhanced, arterial phase, portal phase, and venous phase postcontrast scans showed differences, theredore, for direct comparison, all scans were manually cropped to get the same volumes covering the upper abdomen (Supplementary Fig. 1 of Additional file 1). The three-dimensional volume of the pancreas parenchyma, liver, spleen, iliopsoas muscles, erector spinae muscles, L1 and L2 vertebra, and kidneys were segmented in an artificial intelligence-assisted manner using the TotalSegmentator algorithm [15]. Manual corrections were made to the segmentation masks where necessary. Manual slice-by-slice segmentation of the PCNs was performed by an expert radiologist with over 10 years of experience (Supplementary Fig. 1 of Additional file 1), the area of the PCN was then subtracted from the area of the pancreas parenchyma segmented by the TotalSegmentator algorithm to avoid overlap. In all cases, the segmentation was performed on the original scans, and the same segmentation masks were applied to the corresponding VNC reconstructions since the Liver-VNC algorithm does not change the spatial location of the corresponding voxels. In order to assess the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of the real unenhanced scans as well as the VNC reconstructions, 3 circular ROIs were manually placed to the subcutaneous fat at the middle level of the L3 vertebra on all scans in approximately the same position (Supplementary Fig. 2 of Additional file 1). The size of the ROIs was kept to the maximum, however, it was dependent on the thickness of the subcutaneous fat, therefore the diameter of the circular ROIs was between 9.6 and 31.0 mm.

Quantitative image analysis

During quantitative image analysis, the 70 keV virtual monoenergetic reconstructions of the real unenhanced CT scans were used as reference standards. In this study, the SPP reconstructions of the unenhanced scans that contain all the spectral information were investigated instead of the polyenergetic T3D reconstructions to demonstrate the pure effect of the Liver-VNC algorithm avoiding the influence of the different reconstruction algorithms (SPP vs. T3D) on the results. The reliability of the VNC reconstructions was evaluated by assessing the mean HU density values extracted from the entire volume of the organs. The VNC reconstructions of the real unenhanced scans were also generated to assess the extent to which the algorithm changes the density of organs during iodine subtractions from scans that contain no iodine contrast. The error of the Liver-VNC algorithm (HU_error) was calculated as the difference between the mean density on the VNC reconstruction (HU_VNC) and the mean density on the corresponding real unenhanced phase scan (HU_unenhanced) as proposed by Holz et al. [16].

$$\:{HU}_{error}={HU}_{VNC}-{HU}_{unenhanced}$$

By manually placing 3 circular ROIs in the subcutaneous tissue at the middle level of the L3 vertebra (Supplementary Fig. 2), the mean and the standard deviation (SD) of the HU values were extracted, and the average of the 3 measurements was calculated.

The SNR was defined as the ratio of the mean density and the SD of the voxels of the given organ. The CNR was defined as the difference between the mean density of the given organ and the mean density of the subcutaneous fat, divided by the SD of the voxels of the subcutaneous fat similar to those formulas previously published by Al-Difaie et al. [17] and Si-Mohamed et al. [18].

$$\:SNR={mean}_{tissue}/{SD}_{tissue}$$

$$\:CNR=\frac{{mean}_{tissue}-{mean}_{fat}}{{SD}_{fat}}$$

The calculated SNR and CNR values of the arterial phase VNC (SNR_{VNC[arterial],} CNR_{VNC[arterial]}), portal venous phase VNC (SNR_VNC[portal], CNR_VNC[portal]), venous phase VNC (SNR_VNC[venous], CNR_VNC[venous]), as well as the unenhanced phase VNC (SNR_{VNC[unenhanced],} CNR_{VNC[unenhanced]}) were directly compared to the SNR and CNR values of the real unenhanced scans (SNR_unenhanced, CNR_unenhanced).

For easier interpretability, we summarized the introduced terms of measured HU density values and calculated ratios in Table 1.

Table 1 Abbreviations used for Hounsfield unit density measurements of different scans and reconstructions

The main steps of the data analysis are summarized in Fig. 1. The extracted density values of the organs can be found in Additional file 2.

Subjective reading of CT reconstructions

An expert radiologist with over 10 years of experience performed the side-by-side subjective reading of the unenhanced phase CT scans and the VNC reconstructions. The radiologist was asked to review the presence of pancreatic calcifications, central/mural calcifications of PCNs, and kidney stones, and measure their largest diameter on the real unenhanced scans and VNC reconstructions using the standard bone window. The image quality was then subjectively evaluated by the radiologist based on a 5-point scale; where the 5-point maximum means that the image quality in evaluating the calcification is equal with the real unenhanced phase scan; 4-point means that the image quality is worse but it does not affect the radiological evaluation; 3-point means that the image quality is markedly worse which may affect the radiological evaluation; 2-point means the calcification is barely visible which severely violates the radiological evaluation; and 1-point means the calcification is fully subtracted, not visible. The number of calcified lesions did not allow for detailed statistical evaluation, however, we demonstrate the limitations of the VNC reconstructions in these cases.

Statistical analysis

The density values of the CT scans and VNC reconstructions are reported in HU. During statistical analysis, the Shapiro-Wilk test was used to test normality, followed by Levene’s test for the assessment of the homogeneity of variances between groups. For the comparison of paired groups, either the paired sample t-test or the Wilcoxon signed-rank test was used to compare the mean density values. Correction for multiple testing was applied according to the Benjamini-Hochberg method. A threshold of p < 0.05 was used to decide the statistical significance of all comparisons. The values are reported as mean ± standard deviation. Pearson’s correlation coefficient (r) or Spearman’s rho was used to assess the correlation of HU_error with the contrast enhancement of organs on the postcontrast scans (HU_postcontrast) and the HU_unenhanced. The correlations were classified as weak (r = 0.20–0.39), moderate (r = 0.40–0.59), strong (r = 0.60–0.79), and very strong (r = 0.80-1.00). The reproducibility of the measurements was evaluated based on the mean and 95% confidence interval (CI) values of the intraclass correlation coefficient (ICC) by using a 2-way random effect, single-measurement, absolute-agreement model (ICC 2,1). Bland-Altman plots were used for visualizing the differences between HU_VNC and HU_unenhanced.

The statistical analysis was completed with dedicated packages coded in Python such as “numpy”, “pandas”, “sklearn”, “scipy”, “statsmodels”, “seaborn”, and “matplotlib”.

Results

Patient population

The final patient cohort included 34 patients, of whom 19 were diagnosed with intraductal papillary mucinous neoplasms, 4 with mucinous cystic neoplasms, 4 with serous cystic neoplasms, 3 with adenocarcinomas, 1 with solid pseudopapillary neoplasm, 1 with cystic pancreatic neuroendocrine tumor, 1 with lymphoepithelial tumor, and 1 with unclassified PCN without worrisome features. One patient with serous cystic neoplasm had 2 follow-up CT scans.

The contrast material routinely used in our CT laboratory was changed during the investigated period due to institutional factors; therefore, during contrast-enhanced CT examination, the Ultravist 370 contrast agent was used for 16 patients, while 19 patients received the Iomeron 350 contrast material. Table 2. summarizes the demographic data of the patient cohort.

Table 2 Descriptive data of the patient cohort

Comparison of the mean density values

The violin plots on the differences between the mean densities on the arterial phase (HU_{VNC[arterial]}), portal phase (HU_VNC[portal]), and venous phase (HU_VNC[venous]) VNC reconstructions vs. densities on the real unenhanced scans (HU_unenhanced) can be found in Additional file 1 (Supplementary Figs. 3–6). The correlation plots on the association between HU_VNC and HU_unenhanced as well as the Bland-Altman plots illustrating the association of HU_error with the organs’ average HU density can be found in Additional file 1 (Supplementary Figs. 7–18 and Supplementary Figs. 19–22). All organs had at least one contrast phase that reached a strong correlation except the spleen and kidneys. Moreover, the correlation for the spleen in the portal venous phase did not reach statistical significance.

Comparison of real unenhanced with unenhanced VNC reconstructions

The HU_{VNC[unenhanced]} of the spleen, liver, and paraspinal muscles were significantly (p < 0.0001) higher, while those of the vertebrae were significantly lower (p < 0.0001) compared to the HU_unenhanced values (Table 3). However, the HU_error for the spleen and liver were only 2.35 ± 2.70 and 4.97 ± 2.94 HU, respectively. The HU_{VNC[unenhanced]} values showed very strong correlations with the HU_unenhanced values in all organs except the kidneys (r = 0.78 and 0.73; p < 0.0001) and the spleen (r = 0.45; p = 0.0065) (Table 4).

Table 3 Density differences between VNC reconstructions vs. real unenhanced phase scans, and the correlation with the enhancement on postcontrast phase scans

Table 4 Correlation between the density values of real unenhanced scans and VNC reconstructions

Analysis of the vertebrae

In all postcontrast VNC reconstructions, the most marked HU_error was found in the vertebrae, where the algorithm struggled to differentiate calcium from iodine which resulted in significantly (p < 0.0001) lower density values on the VNC reconstructions (Table 3). The HU_VNC values had very strong correlations with the HU_unenhanced (Table 4). The HU_error also showed very strong correlations with the HU_postcontrast and HU_unenhanced (Table 5).

Table 5 Correlation of the density differences between real unenhanced scans and VNC reconstructions (HU_error) vs (1) densities on the real unenhanced phase scans, and (2) the enhancement on postcontrast phase scans

Analysis of the kidneys

There was no significant difference between HU_unenhanced and HU_{VNC[unenhanced]} in the kidneys. However, the difference between HU_unenhanced and HU_VNC was significant (p < 0.0001) for both kidneys in all postcontrast phases (Table 3) with a HU_error of -4.80 ± 4.05HU for the left and − 4.53 ± 4.75HU for the right kidneys in the arterial phase. The HU_error was markedly higher in the portal phase and even higher in the venous phase (Table 3). Moreover, the HU_VNC values had only a moderate correlation with the HU_unenhanced values (Table 4), and the HU_error also had a moderate association with the HU_arterial and a weak to moderate correlation with the HU_venous (Table 5).

In the arterial phase, 19/35 and 18/35 cases had a HU_error less than ± 5 HU, and an additional 12/35 and 13/35 had a maximum HU_error of ± 10 HU. In the portal venous phase, these were 9/35, 10/35, and 8/35, 6/35, respectively. While in the venous phase, 5/35 and 9/35 cases were within the range of ± 5 HU, and an additional 11/35 and 6/35 had a maximum HU_error of ± 10 HU. The histograms illustrating the number of patients in each category can be found in Additional file 1 (Supplementary Fig. 23).

Analysis of the paraspinal and iliopsoas muscles

The HU_error dominantly resulted in positive values for the paraspinal muscles and the difference between HU_unenhanced and HU_VNC was significant in all phases. Meanwhile, the HU_error of the iliopsoas muscles consistently resulted in negative values but the difference reached significance in only the portal and venous phases (Table 3). The HU_VNC values had very strong correlations with the HU_unenhanced values in both phases for the paraspinal muscles, while the iliopsoas muscles showed strong to very strong correlations (Table 4). The HU_error in the paraspinal muscles showed weak to moderate correlations with the HU_arterial, HU_portal, and HU_venous values, while in the iliopsoas muscles, a significant correlation was only found in the venous phase (Table 5). In both three phases, the HU_error had weak but significant correlations with HU_unenhanced values for all muscles except the left iliopsoas in the venous phase and right iliopsoas in the portal phase (Table 5).

The HU_error was within the range of a maximum of ± 15 HU values in all phases for all muscles. For the paraspinal muscles, 10/35 cases were within the range of ± 5 HU in the arterial phase, 15/35 were within this range in the portal venous phase, while in the venous phase, it was 17/35 cases. The iliopsoas muscles showed better results, 30/35 and 29/35 cases were within the range of ± 5 HU in the arterial phase, 29/35 were within this range in the portal venous phase, while in the venous phase, it was 28/35 and 26/35 cases for the right and left muscles, respectively. The histograms illustrating the number of patients in each category can be found in Additional file 1 (Supplementary Fig. 23).

Analysis of the pancreas and PCNs

The analysis of the pancreas revealed no significant differences between HU_unenhanced and HU_{VNC[arterial]} (Table 3). The HU_VNC values showed very strong correlations with the HU_unenhanced values in all phases (Table 4). Moreover, no correlation was found between HU_error and either HU_unenhanced, HU_arterial, HU_portal, or HU_venous (Table 5).

In the arterial phase, the HU_error values were within the range of ± 5 HU in 29/53 cases, while all but one cases were within the range of ± 10 HU. In the portal venous phase, 20/35 cases had a maximum HU_error of ± 5 HU, while an additional 12/35 were within the range of ± 10 HU. Similar results were found in the venous phase with 26/35 cases within the range of ± 5 HU and an additional 8/35 cases within the range of ± 10 HU. The PCNs showed the best results on the arterial phase, where the HU_error was within the range of ± 5 HU in 18/35 cases, and an additional 11/35 had a maximum HU_error of ± 10 HU. The worst HU_error results were observed in the portal phase, where 7/35 cases were out of the range of ± 15 HU. The histograms illustrating the number of patients in each category can be found in Additional file 1 (Supplementary Fig. 23).

The analysis of the PCNs revealed significant differences between HU_unenhanced and both HU_{VNC[arterial]}, HU_VNC[portal], and HU_VNC[venous], however, the difference in the arterial phase was only − 3.81 ± 5.76 HU (Table 3). The HU_{VNC[arterial]} had a very strong, while the HU_VNC[venous] and the HU_portal−VNC had strong correlations with the HU_unenhanced (Table 4). The HU_error showed a significant correlation only with the HU_arterial values, no correlation was found with HU_unenhanced in either phase (Table 5).

Analysis of the spleen

In the spleen, significant differences were detected between HU_unenhanced and HU_VNC in all phases, although the HU_error was only 1.64 ± 2.66 HU in the arterial phase (Table 3). The HU_VNC had only a moderate correlation with the HU_unenhanced on the arterial and venous phases, while no correlation was found in the portal phase (Table 4). Moreover, the HU_error also had significant, strong correlations with HU_arterial and HU_portal, and a moderate correlation with HU_venous (Table 5).

In the arterial phase, 23/35 cases had a HU_error less than ± 5 HU, and all the cases were within the range of ± 10 HU. In the portal venous phase, 23/35 were within the range of ± 5 HU, an additional 9/35 had a maximum HU_error of ± 10 HU, while the remaining 3/35 had a maximum HU_error of ± 15 HU. In the venous phase, 33/35 cases were within the range of ± 5 HU, and the remaining 2/35 cases had a maximum HU_error of ± 10 HU. The histograms illustrating the number of patients in each category can be found in Additional file 1 (Supplementary Fig. 23).

Analysis of the liver

The liver showed higher HU_{VNC[arterial]} compared to HU_unenhanced, however, the HU_error was only 4.64 ± 2.85 HU. For the portal and venous phases, the HU_error values were even lower, only 2.12 ± 2.60 HU, and 2.50 ± 2.22 HU, respectively. The HU_VNC values showed a very strong correlation with the HU_unenhanced in all phases (Table 4). Meanwhile, the HU_error had no significant correlation with HU_unenhanced in either phase, and it showed only a moderate correlation with HU_portal and a weak correlation with HU_venous (Table 5).

The HU_error values were within a maximum of ± 10 HU in all cases in all contrast phases. In the arterial phase, 17/35 cases had a HU_error less than ± 5 HU, in the portal venous phase, it was 31/35, while the best results were found in the venous phase, where 32/35 cases were within the range of ± 5 HU. The histograms illustrating the number of patients in each category can be found in Additional file 1 (Supplementary Fig. 23).

Intraclass correlation coefficient analysis

The results of ICC analysis showed that most organs had at least one postcontrast phase VNC reconstruction that showed good-to-excellent reproducibility based on the lower value of the 95% confidence interval. Except the vertebrae, spleen, and kidneys which consistently resulted in poor ICC values, and the paraspinal muscles that showed excellent ICC values but with wide confidence intervals. The results of the ICC analysis can be found in Table 6.

Table 6 Interclass correlation coefficient analysis based on the average density values of real unenhanced scans and VNC reconstructions

In the arterial phase, the ICC analysis showed excellent reliability for the pancreatic parenchyma with an ICC of 0.90 [0.80–0.95]. Good reliability was found for PCNs with ICC of 0.76 [0.43–0.89]) and for iliopsoas muscles with ICCs of 0.79 [0.59–0.89], and 0.75 [0.56–0.87]). However, the remaining organs showed wide confidence intervals. In the portal venous phase, only the liver showed excellent reliability with an ICC of 0.93 [0.70–0.97], other organs either had low ICC values or wide 95% confidence intervals. While in the venous phase, only the liver and pancreas showed good reproducibility with ICC of 0.93 [0.46–0.98], and 0.92 [0.42–0.98], respectively.

Bland-Altman plots further supported the results of the ICC analysis (Supplementary Figs. 19–22 of Additional file 1). All cases were distributed around a HU_error close to zero, with almost all measurements within ± 1.96 SD; except for vertebrae, kidneys, and spleen, which showed a huge HU_error, which was significantly affected by the mean of the measurements.

Comparison of image quality

Subjective evaluation of image quality

During the subjective assessment of the 35 CT examinations, the expert radiologist found that the VNC reconstructions were smoother and more blurred compared to the real unenhanced scans. The expert radiologist reported the presence of kidney stones in 10 cases, while, 5 had pancreas calcification, 3 had PCNs with mural calcification, and 2 had PCNs with central calcification. The size of the kidney stones was between 1.9 and 6.9 mm, the pancreatic calcifications evaluated had sizes between 1.9 and 3.5 mm, the mural calcifications had sizes between 2.0 and 2.5 mm, while the two central calcifications had sizes of 2.0 and 6.5 mm on the real unenhanced scans. Comparing the measured diameters of these 20 lesions on the real unenhanced scans vs. VNC reconstructions, the unenhanced phase VNC yielded acceptable results with no significant difference (2.98 ± 1.41 mm vs. 2.48 ± 1.19 mm, p = 0.067), while the arterial, portal, and venous phase VNC reconstructions showed significantly lower measured diameters with 1.64 ± 1.54 mm, 1.65 ± 1.33 mm, and 1.57 ± 1.50 mm each with p < 0.001, respectively.

The expert radiologist rated the image quality on the basis of the radiological assessability of the calcifications according to a 5-point scale. If more than one calcification or more than one kidney stone was presented, the one with the worst assessability on VNC was reported. Therefore, 20 lesions were evaluated in this subanalysis. As a result, the VNC reconstructed from the real unenhanced scans showed acceptable image quality with an average point of 4.3/5.0. From the postcontrast phase VNC reconstructions, the arterial phase had the highest image quality with an average point of 2.65/5.0, followed by the portal venous phase VNC with 2.45/5.0 points on average, while the venous phase VNC showed the worst image quality with average points of 2.4/5.0. A total number of 7/20 lesions resulted in full subtraction in either of the postcontrast phase VNC reconstructions. Although none of the calcifications were fully subtracted from the VNC reconstructions generated from the unenhanced phase scans, the number of fully subtracted calcifications was 6/20 on the arterial phase VNCs, 4/20 on the portal venous phase VNCs, and 7/20 on the venous phase VNCs.

During the subjective reading of the scans, the radiologist found that the Liver-VNC algorithm was less reliable in patients with chronic pancreatitis. During iodine subtraction the algorithm struggled to differentiate calcifications from the iodine contrast material, lowering the density values of small calcifications and blurring their borders (Fig. 2). In kidneys, special attention should be paid to the density values of stones which may also be reduced on the VNC reconstructions (Fig. 3). From the kidneys that showed marked contrast enhancement, the algorithm failed to subtract the iodine contrast which remained detectable in the renal cortex even on VNC reconstructions (Fig. 4). Meanwhile, the algorithm markedly reduced the densities of those kidneys that showed less pronounced contrast enhancement (Fig. 5). Furthermore, in cases where the spleen had marked striking wave-like contrast enhancement, the Liver-VNC algorithm markedly decreased the density values of the highly enhancing areas (Fig. 6).

Fig. 2

Illustration of the limitation of the Liver-VNC algorithm in pancreatic calcifications. Pancreatic calcifications on the unenhanced phase CT scan of a patient with chronic pancreatitis (A). The algorithm blurred the margins and decreased the size of the calcifications (arrow) during the reconstruction of the virtual non-contrast images from the unenhanced phase (B) arterial phase (C), portal phase (D), and venous phase (E) postcontrast CT scans

Fig. 3

Illustration of the limitation of the Liver-VNC algorithm in kidney stones. The difference in the density of a kidney stone between the real unenhanced phase CT series (A) and the virtual non-contrast reconstructions generated from the arterial phase (B), portal phase (C), and venous phase (D) series

Fig. 4

Illustration of the failure of the Liver-VNC algorithm for iodine contrast subtraction from kidneys. The left kidney on an arterial phase CT scan (A), and the difference between the real unenhanced phase CT scans (B) and the virtual non-contrast reconstruction (C)

Fig. 5

Illustration of the excessive decrease in density values during iodine contrast extraction from kidneys by the Liver-VNC algorithm. The difference in the density of the kidney between the real unenhanced phase CT scans (A) and the virtual non-contrast reconstructions of arterial phase (B), portal phase (C), and venous phase (D) scans. The algorithm decreased the mean densities

Fig. 6

Illustration of the failure of the Liver-VNC algorithm for iodine contrast subtraction from the spleen. The striking wave-like contrast enhancement pattern of the spleen on the arterial phase (A) scan. The difference in the density of the spleen on the real unenhanced phase CT scans (B) and the virtual non-contrast reconstructions generated from the arterial phase (C)

Quantitative evaluation of image quality

The subjective impression of smoother images for VNC reconstructions compared to the real unenhanced scans was supported by the results of the quantitative assessment of image noise by calculating the CNR. The CNR_{VNC[unenhanced]} was found to be significantly lower compared to the CNR_unenhanced values for the kidneys (p = 0.016), pancreas (p = 0.041), vertebrae (p < 0.0001), iliopsoas muscles (p = 0.013), and PCNs (p = 0.03). The CNR resulted in significantly lower values on the portal venous phase and venous phase VNC reconstructions compared to the real unenhanced scans for all organs. On the arterial phase, only the vertebrae showed significantly lower CNR_{VNC[arterial]} values (Table 7).

Table 7 Contrast-to-noise ratio of the organs on the virtual non-contrast reconstructions and the real unenhanced CT

The comparison of the SNR between VNC reconstructions and real unenhanced scans showed heterogeneous results for the different organs on different phases (Table 8). The SNR_{VNC[unenhanced]} were significantly higher for the spleen (p < 0.0001), kidneys (p < 0.0001, p < 0.0001), liver (p < 0.0001), pancreas (p < 0.0001), paraspinal muscles (p < 0.0001), and iliopsoas muscles (p < 0.0001) compared to the SNR_unenhanced, while no significant difference was found in the vertebrae and PCNs. For all postcontrast phases, the muscles showed significantly higher SNR_VNC values, while the kidneys and PCNs had significantly lower SNR_VNC values compared to SNR_unenhanced.

Table 8 Signal-to-noise ratio of the organs on the virtual non-contrast reconstructions and the real unenhanced CT

Discussion

In our study, we analyzed 12 abdominal organs to investigate whether VNC images reconstructed from arterial, portal, and venous phase spectral CT scans using the Liver-VNC algorithm can replace real unenhanced scans in abdominal imaging., moreover, we directly compared VNC reconstructions of unenhanced phase scans with real unenhanced scans. The difference in density values between the VNC and real unenhanced scans was significant for all postcontrast scans in the kidneys, liver, vertebrae, paraspinal muscles, and PCNs. However, the HU_error was only 2.12 ± 2.60 HU and 2.5 ± 2.22 HU for the liver on the portal venous and venous phases, respectively. The density values of the VNC reconstructions also showed a strong to very strong correlation with those on the real unenhanced scans in all investigated organs except the spleen, and kidneys. Finally, we also highlighted the most common pitfalls of VNC reconstructions including the subtraction of kidney stones, pancreatic calcifications, PCN calcifications; and the impaired iodine contrast subtraction from the kidneys and spleen that can limit the introduction of the algorithm to daily clinical practice. Hence, this study contributes to the emerging body of literature evaluating the feasibility and potential of VNC reconstructions of PCD-CT examinations.

By the visual inspection of the reconstructed images, we found that the VNC reconstructions had a more blurred appearance compared to the real unenhanced scans which is supported by the results of the quantitative image noise assessment analysis, as the CNR proved to be significantly lower for all organs in all phases in accordance with the results of existing literature [10, 17].

Previous publications quantitatively validating the reliability of VNC reconstructions mainly focused on the comparison of average HU density values of the investigated organs. Niehoff et al. [11] tested the prior version VNC algorithm of the same PCD-CT model. The authors reported significant differences between HU_VNC and HU_unanhanced in both the arterial and venous phase scans for all the investigated organs. In our study, we confirmed significant differences between HU_VNC and HU_unanhanced in the spleen in the arterial and portal phases, and in the paraspinal muscles, kidneys, liver, and vertebrae in all three postcontrast phases (Table 3). However, in our study, the HU_VNC of paraspinal muscles and the liver showed a markedly stronger correlation with HU_unanhanced (r > 0.90) on all postcontrast phases. These differences can be explained by the fact that we used the Liver-VNC algorithm of the vendor instead of its previous VNC algorithm, therefore our results suggest that this software version may be more reliable.

In their recently published study, Schoenbeck et al. [12] compared the performance of the prior VNC algorithm and the later Liver-VNC algorithm in various abdominal organs. This study was based on the same PCD-CT model of the vendor that we used. The authors found significant differences between HU_{VNC[unenhanced]} and HU_unenhanced in the liver, renal cortex, aorta, paraspinal muscle, and subcutaneous fat, but reported no significant differences in the spleen using the VNC algorithm. On the contrary, in our study, by using the Liver-VNC algorithm we found no significant differences in the kidneys, pancreas, iliopsoas muscles, and PCNs, however, we confirmed significant differences in the liver and paraspinal muscles, and we also reported significant differences in the spleen and vertebrae. The authors found significant differences between HU_VNC[portal] and HU_unenhanced, whether the earlier VNC or the later Liver-VNC algorithm was used in all organs except the paraspinal muscles. In our study, we confirmed the significant difference in all organs in the portal venous phase and that the algorithm decreases the density of the kidneys and increases the density of the liver. Moreover, we also investigated the correlation of the HU_error with HU_unenhanced and HU_postcontrast values.

Interestingly, a previous study on a dual-energy CT scanner by Lin et al. [8]. investigating the kidneys obtained contrary results. They found similar cortical HU_VNC values in the corticomedullary phase and significantly increased cortical HU_VNC in the nephrographic phase. Although the average HU_error values that Lin et al. reported on the corticomedullary phase were only − 0.4 HU and 3.1 HU for the cortex and medulla, they had wide ranges of -21.7 − 22.7 HU and − 30.0 − 21.2 HU. This discrepancy can be at least partially explained by the difference in contrast injection timing and flow rates which affects the HU_error values as we demonstrated by correlating the HU_error with the HU_postcontrast values.

The ranges of HU_error in our study were categorized as ± 5 HU, ± 5–10 HU, ± 10–15 HU, ± 15–20 HU, and > 20 HU. Comparing these results to those published by Çamlıdağ on VNC reconstructions of dual-energy CTs on the nephrographic phase [19], slightly better results were found in our study for the liver and spleen, comparable results for the pancreas, and slightly worse results for the kidneys. The authors reported that 25/142 and 4/142 of the cases had density differences of > 10 HU and > 20HU for the liver, while in our study, none of the cases had HU_error larger than ± 10 HU. This difference may be explained with that we used the Liver-VNC algorithm which takes the fat into account during material decomposition. For the spleen, the authors reported that 27/142 and 2/142 cases had density differences of > 10 HU and > 20 HU, while in our study, all the cases were within the range of ± 10HU on the arterial and venous phases, and only 3/35 cases had HU_error larger than ± 10 HU on the portal venous phase. For the kidneys, approximately half of the cases showed a HU_error larger than ± 10HU in our study, while Çamlıdağ reported that only 24/142 and 34/142 patients had density difference > 10 HU for the left and right kidneys, respectively.

We consider the observed HU_error acceptable for the liver, as in daily clinical practice, the CT-based rough assessment of the presence of significant hepatic steatosis relies on the average density measurement on the unenhanced phase scans with a threshold of < 40 HU [20]. An average density measurement error of approximately 2–4 HU that we reported may not increase the false positive findings of hepatic steatosis, however, our current research did not aim to cover the detailed reliability analysis of the algorithm in fatty liver disease, therefore, further investigation is needed in this field.

The density measurement of the pancreatic parenchyma on unenhanced CTs can help radiologists diagnose pancreatic lipomatosis. As the HU_error was only − 1.47 ± 4.87 HU for the arterial phase, -4.9 ± 5.9 HU for the portal phase, and − 3.33 ± 2.95 HU for the venous phase VNC reconstructions in our study, we suggest that it would not increase the false positive findings of pancreatic lipomatosis. Similarly, the average density measurement of the paraspinal muscles on unenhanced CT has a role in the diagnostics of fatty infiltration of the muscles, however in our study, the Liver-VNC algorithm tended to increase the density values of the paraspinal muscles.

The measurement of lower average density values of the kidney parenchyma on unenhanced CT scans, the so-called “pale kidney sign” is an additional finding that can suggest renal edema and urinary obstruction. Therefore the dramatic decrease of average density values on the VNC reconstructions that we reported may affect the diagnosis. Moreover, the inappropriate contrast-subtraction can cover small renal stones resulting in false negative findings.

The assessment of calcifications on unenhanced phase CT scans is crucial in the diagnostics of pancreatic cystic neoplasms [13], moreover reading the unenhanced phase CT scans is also critical in the diagnostics of chronic calcifying pancreatitis [21], and the characterization of kidney stones [22]. Therefore, our study had a special focus on the semi-quantitative analysis of VNC reconstructions’ image quality regarding the radiological evaluation of such small calcifications. The iodine-based contrast media and calcium-containing structures (vertebrae, calcifications) have low differences in spectral attenuation, and the k-edge of iodine and calcium are close to each other [23]. Therefore, the Liver-VNC algorithm cannot differentiate between the two perfectly during material decomposition which is a well-known limitation. For a more detailed description of the k-edge imaging and its limitations, we kindly refer the readers to the recently published methodological paper of Jost et al. [23]. This limitation had the most noticeable effect on the vertebrae in our study which resulted in significantly lower HU values on the VNC reconstructions compared to real unenhanced scans, which can affect the assessment of CT-based osteoporosis. Due to this limitation, the algorithm was less reliable in the assessment of central or mural calcifications of PCNs, and in patients with chronic pancreatitis, where it struggled to differentiate calcified lesions from iodine, lowering the density values of small calcifications, blurring their borders, or even fully subtracting them (Fig. 2) resulting in a clearly inferior image quality during semi-quantitative assessment. This finding complies with those reported by Mileto et al., who assessed the image quality of VNC reconstructions of a dual-energy scanner in pancreas imaging and also reported the partial subtraction of small pancreatic calcifications in 2 cases [24]. We also observed that in the case of high-density calcium stones in the kidney, the Liver-VNC algorithm markedly lowered the density values of the stones (Fig. 3) which can have therapeutic consequences. This observed phenomenon is in accordance with the findings of Dodig et al. [25] and Gezer et al. [26] who reported accuracies of 72.41–80.46% [25] and a sensitivity of 66.7% [26] for VNC reconstructions in the detection of kidney stones when compared to the real unenhanced scans.

The inaccuracies of the algorithm discussed above can be a major limitation of the application in pancreas imaging and kidney stone assessment. To overcome this limitation, the two main directions are improving the current VNC algorithms designed for iodine-based contrast-enhanced images, and the development of alternative contrast materials. On one hand, the introduction of a VNC algorithm specifically designed to preserve calcifications in parenchymal organs and the development of a VNC algorithm optimized for organs showing marked contrast-enhancement (e.g. kidney, spleen) would be important future research directions. On the other hand, the usage of iodine-based contrast materials is a severe limitation of spectral imaging-based material decomposition due to their suboptimal spectral attenuation range. Therefore, the field of developing new, alternative contrast materials also gained increasing attention in the literature [23]. A phantom study by Amato et al. [27] demonstrated that contrast agents with higher atomic numbers such as gadolinium could serve as a better alternative for PCD-CT. Moreover, Kim et al. [28] also proved that the reconstructed contrast maps successfully differentiated gadolinium from iodine and calcium. These phantom studies demonstrate that the introduction of new imaging protocols using alternative contrast agents could successfully overcome the limitations of calcium subtraction in the reconstruction of VNC images and contrast agent maps in the future.

Our study has some limitations that must be taken into consideration. Our study had a retrospective study design with a low number of cases including a special population of patients with PCNs. The evaluation of the reliability of VNC reconstructions in special pathologies of the given abdominal organs e.g. different stages of liver fibrosis in the case of the liver, degree of sarcopenia in the case of the muscles, or chronic renal failure in the case of the kidneys were out of the scope of our current research. In our study, the small number of patients did not allow us to investigate the effect of contrast agent concentration and contrast agent injection flow rate on the reliability of the Liver-VNC algorithm. Our study was based on the three-dimensional volumetric assessment of organs focusing on the changes in average HU density values caused by the Liver-VNC algorithm that affect the entire organ. This method, however, cannot capture fine differences between real unenhanced scans and VNC reconstructions such as the subtraction of small-size calcifications and kidney stones. The limitation of the Liver-VNC algorithm in these cases was observed by an expert radiologist during the subjective reading of real unenhanced scans and VNC reconstructions. The number of patients with these lesions in our study did not allow statistical analysis, but the limitations of VNC reconstructions in these cases are important and may prevent the widespread clinical use of VNC. Therefore, further studies based on larger patient populations are needed in this field.

Conclusions

In conclusion, our study confirmed the results of previous literature in the case of the liver parenchyma analysis, as we demonstrated the reliability of mean density measurement on VNC reconstruction especially in the portal and venous phases. The results of our study also demonstrated that the degree of enhancement on postcontrast scans as well as the base enhancement on the unenhanced scans affect the reliability of VNC reconstructions for several organs. Although mean density measurements on VNC reconstructions were found to be reliable in at least one postcontrast phase for most organs, further improvements are needed before the VNC reconstructions can be introduced to clinical applications in abdominal imaging and can be utilized to examine the spleen, kidneys, and vertebrae.