Dual-source computed tomography of the lung with spectral shaping and advanced iterative reconstruction: potential for maximum radiation dose reduction

A total of 64 patients with dual-source CT examinations of the lung were enrolled in this study. They were retrospectively selected from four examination protocols available in our department. In 16 patients (age 11.2±5.0 years, median 12.4 years, range 2.9–17.7 years) a full-dose (FD) dual-source CT of the lung had been conducted. Forty-eight other patients had been examined using one of three reduced-dose protocols with tin prefiltration (Sn) established in our department (Sn96: n=16, 10.3±6.1 years, median 11.6 years, range 1.3–17.7 years; Sn64: n=16, 13.1±3.4 years, median 13.1 years, range 5.6–18.0 years; Sn32: n=16, 11.4±4.2 years, median 12.8 years, range 4.8–17.6 years). The Sn protocols had been implemented at our institute in order to gradually reduce radiation exposure in clinical routine. Patients of the different groups were matched for age, weight and body mass index. Among all groups there was no significant difference in patient characteristics (Table 1).

All patients had been referred for CT to further investigate suspected or known non-cancer lung diseases such as cystic fibrosis, primary ciliary dyskinesia, prolonged course of pneumonia, chronic lung complications of pneumonia, suspected pulmonary hemorrhage, aspiration pneumonitis, pulmonary Langerhans cell histiocytosis, tuberculosis, and atelectasis or pleural effusion of unclear origin.

All dual-source CT examinations were performed using the same third-generation scanner (Somatom Definition Force, Siemens Healthcare). CT parameters were as follows: 0.25 s gantry rotation time, detector collimation of 2x96x0.6 mm, slice collimation of 192×0.6 mm using z-flying focal spot technique, spiral pitch factor 3.0, tube voltage modulation switched off. In the full-dose protocol, patients were examined at a 100-kV setting with automatic exposure control (reference tube current time product per rotation 64 mAs; CareDose4D, Siemens Healthcare). In all other protocols 0.6-mm tin prefiltration was applied. Because tin prefiltration is only available at 100-kV and 150-kV tube voltage, with higher diagnostic dose efficiency at 100 kV [9], the lower kV setting is used in our department. For the three examination protocols with spectral shaping, default values of reference tube current–time product per rotation were 96 mAs/64 mAs/32 mAs. Examinations were performed in supine position with elevated arms from the upper to the lower thoracic aperture. If necessary, a body-weight-adapted dose of iodinated contrast medium was injected intravenously (iomeprol 300 mg/mL, Iomeron, Bracco Imaging, Konstanz, Germany; or Accutron CT-D, Medtron AG, Saarbrücken, Germany).

Primary image data were automatically generated with a slice thickness of 0.6 mm using filtered back-projection (FBP). Additionally, all data sets of examination protocols including tin prefiltration were generated with advanced iterative reconstruction utilizing a medium, an intermediate and a strong increment (ADMIRE strengths 2/3/4). Slice thickness was 0.6 mm, in these protocols, too. In our clinical practice we observed adequate diagnostic quality on full-dose examinations when ADMIRE 2 was used. Consequently, reconstruction of ADMIRE 3 and ADMIRE 4 had not been performed at the time of examinations and thus was not available in the retrospective setting of this study. Iterative reconstruction is characterized by repeated forward and back projection of raw data and image data in combination with statistical modeling. The repeated comparison of projected raw data with the measured data allows removal of geometric imperfections. ADMIRE is built upon these principles, with substantial modifications, allowing a high iteration speed [7]. It has been shown that ADMIRE has the potential to significantly improve image quality while reducing noise and artifacts in CT scans [8, 10, 11]. In ADMIRE, images are reconstructed by minimizing the objective function incorporated with an accurate system model, a statistical noise model, and a prior model [12].

All images were anonymized and transferred to a post-processing 3-D console (SyngoVia VA30A; Siemens Healthcare).

Images were analyzed independently by two radiologists (O.R. and M.H., with 25 years and 10 years of experience in pediatric lung CT, respectively), following the European Guidelines on Quality Criteria for CT. The ratings of the two readers were averaged. For all images, a dedicated lung convolution kernel (Bl57) was used, as recommended by the manufacturer. Images were interpreted in axial, coronal and sagittal orientation with 1-mm slice thickness using a multiplanar imaging tool (MM Reading, SyngoVia VA30A; Siemens Healthcare). Maximum- and minimum-intensity projections were allowed to be used at the discretion of the readers. The default window setting was center –600 HU and width 1,700 HU and could be individually adjusted by the readers.

We rated diagnostic confidence as well as detectability of the following anatomical structures on a 4-point Likert scale (1 unacceptable, 2 acceptable under limited conditions, 3 probably acceptable, 4 fully acceptable): medium-size and small pulmonary vessels, tertiary bronchi, lung fissures, lung parenchyma. We also rated suspicious lung lesions with respect to detectability, contrast and contour sharpness using the same 4-point Likert scale.

To assess image quality, we measured noise in the tracheal lumen on 1.0-mm-thick axial images of all datasets (FBP, ADMIRE 2/3/4). Ten randomly selected patients were evaluated ex ante to detect the optimal surface of the circular region of interest (ROI) with respect to the anatomical target regions. Thus, the defined size of ROI was 0.4 cm² for older children and adolescents and 0.2 cm² for smaller children. For each axial image, we performed and averaged three measurements. Image noise was defined as the standard deviation of the attenuation value.

Radiation exposure was assessed as volumetric CT dose index (CTDI_vol) and dose–length product (DLP). Estimated effective dose (ED) was calculated as DLP·k, using an individual linear interpolation of the conversion factor reported in literature for chest CT at 100 kV between neonates (k₀=0.0739 mSv/mGy·cm), 1-year-olds (k₁=0.048 mSv/mGy·cm), 5-year-olds (k₅=0.0322 mSv/mGy·cm), 10-year-olds (k₁₀=0.0235 mSv/mGy·cm) and 18-year-olds (k₁₈=0.0144 mSv/mGy·cm) as a function of days of age [8, 13].

Statistical analysis was performed using SPSS software version 25 (IBM, Armonk, NY) and WINPEPI (Abramson JH, Hebrew University, Jerusalem). Values are given as mean ± standard deviation if normal distribution was assumed by Kolmogorov–Smirnov tests. Nominal variables were also expressed as frequencies. For multiple comparisons one-way analysis of variance (ANOVA) multiple comparison test with Bonferroni and Games–Howell post hoc pairwise comparisons were applied. All tests were performed two-sided, and P<0.05 was considered to be statistically significant. We calculated proportion of inter-rater disagreement and information-based measure of disagreement (IBMD). IBMD measures the level of disagreement between two or more observers. A value of 0 indicates no disagreement, whereas a value of 1 indicates total disagreement [14].

Diagnostic confidence improved in all Sn groups with increasing strength levels of ADMIRE (Table 2). There was no significant difference in diagnostic confidence between the full-dose group reconstructed with ADMIRE 2 and the Sn96 group reconstructed with ADMIRE 4 (FD_ADM2 vs_. Sn96_ADM4: P=0.092). Although differences between the FD_ADM2 group and the Sn64_ADM4 and Sn32_ADM4 groups were significant (FD_ADM2 vs_. Sn64_ADM4: P=0.008; FD_ADM2 vs_. Sn32_ADM4: P<0.001), diagnostic confidence reached a Likert score >3 in the Sn64_ADM4 group. This was not true for the Sn32_ADM4 group (2.7). For further information see Table 2. The two readers disagreed in 50 of 224 ratings (22%, IBMD 0.10, 95% confidence interval [CI] 0.08–0.12).

In all Sn groups, detectability of anatomical structures improved with increasing strength levels of ADMIRE (Table 2). Compared to FD_ADM2, values for Sn96_ADM4, Sn64_ADM4 and Sn32_ADM4 were 3.4±0.6 vs. 3.2±0.6, 3.1±0.4 and 2.5±0.6 for small vessels; 3.8±0.4 vs. 3.8±0.4, 3.7±0.4 and 3.5±0.5 for tertiary bronchi; and 3.5±0.7 vs. 2.8±0.7, 3.0±0.6 and 2.2±0.5 for lung fissures. Except for lung parenchyma, no significant differences in detectability of anatomical structures were found between FD_AM2 and Sn64_ADM4. On the other hand, differences between Sn64_ADM4 and Sn32_ADM4 were statistically significant with the exception of tertiary bronchi and lung parenchyma. More detailed information regarding all evaluated anatomical structures and corresponding values of significance is described in Table 2. An example is given in Fig. 1. The two readers disagreed in 281 of 1,120 ratings (25%, IBMD 0.10, 95% CI 0.08–0.12).

A total of 231 lung lesions were identified on the CT examinations. Mean diameters of the lesions were 6.1±5.6 mm in the full-dose, 5.0±4.5 mm in the Sn96, 6.2±5.9 mm in the Sn64, and 7.6±6.6 mm in the Sn32 groups (ANOVA: P=0.095; Table 3). The lesions comprised subpleural, peribronchovascular or centrilobular nodules, mucoid impaction, tree-in-bud opacities, septal thickening, local ground-glass opacity, circumscribed consolidations, abscess formation, bronchiectasis, pneumatoceles and cavitations.

Concerning detectability, contrast and contour sharpness of lesions, differences between the FD_ADM2 group and all Sn_ADM4 groups turned out to be significant in terms of statistics (P<0.001; Table 3). Nevertheless, a Likert score value clearly >3 was reached in the Sn96_ADM4 and Sn64_ADM4 groups regarding lesion detectability (3.4 and 3.3; Fig. 2). Moreover, Sn64_ADM4 only marginally missed a score value of 3 points concerning contrast (2.9) and contour sharpness (2.8). Compared to the Sn64_ADM4 group, there were lower score values in the Sn32_ADM4 group, being significant concerning contour sharpness (2.3, P<0.001; Table 3). An example is given in Fig. 3. The two readers disagreed in 648 of 2,454 ratings (26%, IBMD 0.12, 95% CI 0.10–0.13).

Regarding tin prefiltration, lowering of the reference mAs effected an increase of noise. With increasing strength levels of ADMIRE, noise significantly decreased in all study groups (Fig. 4). Corresponding values for FBP and ADMIRE 2, 3 and 4 were 78.3±12.1 HU, 61.3±10.5 HU, 50.3±9.4 HU and 40.6±8.0 HU in the Sn96 group; 92.3±18.5 HU, 67.3±12.9 HU, 57.1±11.6 HU and 45.7±9.3 HU in the Sn64 group; and 120.9±15.4 HU, 90.2±12.7 HU, 75.7±11.0 HU and 61.5±9.3 HU in the Sn32 group (P<0.001). An example is given in Fig. 5. Noise value of the Sn64_ADM4 group did not statistically differ from that in the FD_ADM2 group (45.7 vs. 38.8 HU, P=0.132; Fig. 4). On the other hand, noise was significantly higher in the Sn32_ADM4 group compared to the FD_ADM2 group (61.5 vs. 38.8 HU; P<0.001) and even to the Sn64_ADM4 group (61.5 vs. 45.7 HU; P<0.001).

Compared to the full-dose group, the use of tin prefiltration led to a significantly lower radiation exposure in the Sn96, Sn64 and Sn32 groups. Mean CTDI_vol was 2.17±1.2 mGy vs. 0.31±0.14 in the Sn96 group, 0.24±0.10 mGy in Sn64, and 0.13±0.09 mGy in Sn32 (full dose vs. Sn_all: P<0.001). Corresponding mean DLP was 70.3±44.0 mGy·cm in the FD group vs. 10.5±5.7 mGy·cm in the Sn96 group, 8.2±3.4 mGy·cm in Sn64, and 3.9±2.7 mGy·cm in Sn32 (P<0.001). Consequently, this led to a mean effective dose (ED) of 1.26±0.54 mSv in the FD group vs. 0.21±0.07 mSv in the Sn96 group, 0.14±0.05 mSv in Sn64, and 0.07±0.04 mSv in Sn32 (P<0.001). Among others, dose reduction was statistically significant between full-dose and Sn64 groups, as well as between Sn64 and Sn32 groups (P<0.001). Effective dose was reduced to 16.7%, 11.1% and 5.5% in the Sn96, Sn64 and Sn32 groups, respectively, compared to the full-dose group (Table 4).