INTRODUCTION

Computed tomography (CT) is the current modality of choice for evaluation of the morphological features of lung nodules due to its superior spatial resolution among all imaging modalities.[1] Assessment of lung nodule features on CT may indicate the likelihood of malignancy[2,3] and is the basis of evaluation systems which provide management recommendations.[1,4-7] New CT technologies have recently emerged into the clinical setting, namely ultra-high-resolution CT[8,9] (UHRCT) and photon counting CT (PCCT).[10,11] Both technologies have achieved improvements in spatial resolution relative to conventional CT.[11,12] which offer the potential for improved clinical decisions on how to treat patients.[13-16]

The Aquilion Precision CT scanner (Canon Medical Systems, Otawara, Japan) is an UHRCT scanner featuring a 0.25 × 0.25 mm detector element size, 0.4 × 0.5 mm focal spot size, and is capable of resolving objects as small as 150 µm.[12] The Precision UHRCT scanner has been utilized clinically to evaluate coronary artery stenosis,[14] invasiveness of lung adenocarcinoma,[15] pulmonary calcifications,[17] and temporal bones.[18] Tsubamoto et al.[19] clinically evaluated UHRCT for the visualization of pulmonary nodules, but they did not compare it to conventional resolution CT likely because of the inherent ethical issue of performing a repeat scan on the same patient. To address this limitation with previous studies, we utilize a validated simulation algorithm for generating normal-resolution (NR) reconstructions (Synthesized normal resolution image [NRsim][20]) as a representation of images acquired on a conventional CT scanner. This methodology facilitates the comparison of two different technologies on the exact same patient. The purpose of this work is, therefore, to evaluate the efficacy of UHRCT in comparison to conventional-resolution CT for the characterization of lung nodules on the same patient.

MATERIAL AND METHODS Patient population

The Institutional Review Board at UC Davis approved this retrospective study with a waiver for informed consent. Details regarding patient selection are shown in Figure 1. The final analysis dataset consisted of 104 patients.

Figure 1: Flow chart detailing patient selection. (UHRCT: Ultra-high-resolution CT).

Export to PPT

UHR imaging

All scanning was performed on the UHRCT system in super high-resolution (SHR) mode with acquisition and reconstruction parameters as indicated in Table 1. The raw projection data from the scanner were stored for all exams included in the study. For ease of discussion and to minimize vendor-specific terminology, all images reconstructed using SHR acquisitions on the UHRCT scanner will hereafter be referred to as “UHR.”

Table 1: CT acquisition and reconstruction parameters used in this study.

Parameter Value Resolution mode (detector size) SHR (0.25 mm×0.25 mm) Detector configuration (collimation width) 0.25×160 (40 mm) Helical pitch 0.806 Rotation time (s) 0.35 Tube voltage (kV) 120 Focal spot size* (mm) (# cases) 0.6×0.5 (n=17)
0.6×1.3 (n=87) TCM: Average mA (min, max) 360 (95, 500) Body CTDIvol: Average
mGy (min, max) 11.2 (1.9, 19.4) Display field-of-view (mm) 350 Reconstruction algorithm; kernel AIDR3D STD; FC52 (lung) UHR: Reconstructed voxel size (mm) 0.3125×0.3125×0.25 NRsim: Reconstructed voxel size (mm) 0.6250×0.625×0.50 Synthesizing NR images

A validated algorithm for synthesizing NR images from high-resolution acquisitions (NRsim[20]) was utilized to generate corresponding NR images for all 104 cases included in this study. NRsim images were reconstructed with a voxel size twice that of the corresponding UHR images in each dimension [Table 1]. All other reconstruction parameters were the same. For ease in discussion, NRsim images are hereafter referred to as “NR.”

Reader study design

An experienced thoracic radiologist used all images on our picture archiving and communication system (PACS) to record the location and density (ground glass, part-solid, and solid) of one lung nodule for each patient included in the study. All cases were viewed and scored using custom visualization software developed in-house using Matlab R2022 (Mathworks, Natick, Mass) on a high-performance workstation [Figure 2]. A training session, consisting of 3 cases not included in the study, was included for all readers to familiarize themselves with the software.

Image scoring and analysis

Five radiologists, 3 board-certified with 16, 14, and 9 years of experience and 2 residents with 4 and 2 years of experience, blinded to the acquisitions and reconstruction parameters, participated in this study. An independent reading approach[21] was utilized, meaning that SHR and normal resolution (NR) images from the same case were shown separately. SHR and NR images for the 104 cases were randomly assigned into two reading phases, with no cases repeated in the same phase. These phases were separated by a minimum 4-week washout period to reduce memory bias. To quantify intrareader agreement, a third reading phase was performed whereby 4 radiologists read 30 cases randomly selected from the cases included in the study, with an equal split between UHR (n = 15) and NR (n = 15).

The readers were instructed to score the overall image quality (IQ) throughout the entire lung parenchyma on a 3-point Likert scale (1 = poor IQ; 2 = adequate; and 3 = excellent) and the nodule margin clarity on a 5-point Likert scale (1 = not visible; 2 = poor visualization/delineation; 3 = adequate; 4 = good; and 5 = excellent). Readers were also instructed to define the nodule density (ground glass, part solid, and solid) and their confidence in that density classification on a scale from 0% to 100% in 1% intervals. Finally, the readers were instructed to measure the long and short axes of the nodule in the axial plane [Figure 2].

Figure 2: Screenshot of the visualization and annotation software used in the reader study. The software reads in the DICOM volume data sets and displays three orthogonal views as shown. A display panel was used to capture reader scores and annotations for the images. The software allows navigation through all planes simultaneously, zoom, pan, and adjustment of window/level settings. Each nodule is automatically indicated on the software with a pink ring as shown.

Export to PPT

Image noise was quantified at the aortic arch level within the trachea, a location chosen due to its predominantly homogenous composition of air. Image noise, calculated as the standard deviation of all voxels, was measured within a 1-cm diameter spherical volume-of-interest, with its center location determined by an experienced thoracic radiologist.

Statistical analysis

All statistical analysis was performed using Matlab R2022. Differences in all metrics between UHR and NR were tested using the Wilcoxon signed-rank test, with statistical significance indicated by P < 0.05. For margin clarity and density confidence scores, sub-analyses were performed for each of the three categories of nodule density (ground glass, part solid, and solid) which was determined by an experienced thoracic radiologist (section 2.4). When using the statistical test to compare scores, each reader’s score was averaged per case, and the test was applied to these average values. For the long/short axis measurements, the standard deviation was computed across the five readers for each case and used to assess the significance of differences between UHR and NR.

Kendall’s τ coefficient was used to measure the ordinal association between margin clarity and IQ. This test is useful for determining if there is a consistent ranking pattern of these metrics by readers across all nodules. Coefficient classes were defined using the following criteria: Poor agreement (τ < 0.20), fair (0.20 ≤ τ < 0.40), moderate (0.41 ≤ τ < 0.60), good (0.61 ≤ τ < 0.80), and very good (τ ≥ 0.81).[22] Intrareader agreement for margin clarity, IQ, and short/long axes dimensions was quantified using the intraclass correlation coefficient (ICC).[23] For this work, ICC values <0.5 are defined as poor reliability, values between 0.5 and 0.75 as moderate, values between 0.75 and 0.90 as good, and values >0.90 are defined as excellent reliability. Inter-reader agreement for margin clarity and IQ was quantified using Fleiss’ kappa (k) statistic to calculate the agreement in classification over that which would be expected by chance. k value classes were defined using the following criteria: poor agreement (k < 0), slight (0 ≤ k ≤ 0.2), fair (0.2 < k ≤ 0.4), moderate (0.4 < k ≤ 0.6), substantial (0.6 < k ≤ 0.8), and perfect (0.8 < k ≤ 1).

RESULTS

The patient cohort included in this study consisted of 57 females and 47 males. The mean age of the patients was 66 years, with a standard deviation of 13.2 years. Most of the nodules were classified as solid (n = 67), whereas 21 were part solid and 16 were ground glass.

Figure 3a details the margin clarity score results separated for the three different nodule types. Margin clarity scores were significantly higher for UHR compared with NR for ground glass, part solid, and solid nodules (P < 0.01 for all cases). In addition, margin clarity scores were generally the lowest for ground-glass nodules and the highest for solid nodules. When results were pooled across all nodule densities, UHR resulted in a significantly higher margin clarity score than NR (P < 0.001) as shown in Figure 3b. UHR acquisitions resulted in significantly higher IQ scores compared to NR (P < 0.001) as shown in Figure 4. Kendall’s correlation coefficient of margin clarity and IQ showed moderate agreement (τ = 0.566) for UHR and good agreement (τ = 0.637) for NR. Intrareader agreement was poor for both margin clarity (ICC = 0.017) and IQ (ICC = 0.069) scores. Inter-reader agreement for margin clarity was slight (k = 0.116) for UHR and slight (k = 0.011) for NR. For IQ scores, intrareader agreement was slight (k = 0.017) for UHR and poor (k = −0.130) for NR.

Figure 3: (a) Bar plot of mean margin clarity across all readers for normal resolution and ultra-high-resolution plotted separately for each nodule density and (b) box plots of the average across all nodule densities. The errors bars in (a) represent the 95% confidence interval across all cases. The red line in the middle of each box in (b) is the sample median, and the blue bottom and top of each box are the 25th and 75th percentiles of the sample. Statistically significant differences (P < 0.05) are indicated by an asterisk (*). (NR: Normal resolution, UHR: Ultra-high-resolution, GG: Ground glass, PS: Part solid, S: Solid).

Export to PPT

Figure 4: Box plot of image quality scores averaged across all readers and compared between normal-resolution and ultra-high-resolution. Statistically significant differences (P < 0.05) are indicated by an asterisk (*). NR: Normal resolution, UHR: Ultra-high-resolution. The red line in the middle of each box is the sample median, the blue bottom and top of each box are the 25th and 75th percentiles of the sample, and the red “+” signs are outliers.

Export to PPT

Results for the density confidence scores are shown in Figure 5. Overall, the confidence in assigning nodule densities was significantly higher for UHR in comparison to NR for ground glass, part solid, and solid nodules (P < 0.01 for all cases). When results were pooled across all nodule densities, UHR resulted in a significantly higher density confidence score compared to NR (P < 0.001), as shown in Figure 5a and b. Intra-reader agreement was poor for nodule density confidence scores (ICC = 0.223). Figure 6 provides examples of lung nodules on NR and UHR with noticeable differences in nodule density [Figures 6a-b] and margin clarity [Figures 6c-d].

Figure 5: (a) Bar plot of mean density confidence scores across all readers for normal resolution and ultra-high-resolution plotted separately for each nodule density and (b) box plots of the average across all nodule densities. The error bars in (a) represent the 95% confidence interval across all cases Statistically significant differences (P < 0.05) are indicated by an asterisk (*). (NR: Normal resolution, UHR: Ultra-high-resolution, GG: Ground glass, PS: Part solid, S: Solid).

Export to PPT

Figure 6: Image examples of lung nodules (purple circles) from (a,b) a 61 year old male and (c,d) a 83 year old female on normal resolution (NR) and ultra-high-resolution (UHR) CT. Lung nodule margin is less clearly visualized for (a) NR (b) compared with UHR and (c) groundglass density determination is less confidently made for NR (d) compared with UHR.

Export to PPT

Table 2 details the results of the variability for the short and long axis measurements for each nodule, quantified using standard deviation. While overall variability was higher for NR, no statistically significant differences were observed. Intra-reader agreement was moderate for short-axis measurements (ICC = 0.731) and good for long-axis measurements (ICC = 0.807).

Table 2: Mean standard deviation in measurement of nodule short and long axes across all 5 readers. P-values were computed using the Wilcoxon signed-rank test.

Mean standard deviation NR UHR P-value Nodule short axis (mm2) 1.053 0.951 0.1120 Nodule long axis (mm2) 1.179 1.100 0.2303

Image noise, measured in the trachea at the aortic arch level, was significantly higher for UHR as shown in Figure 7a. Averaged across all cases, image noise was 57.7 Hounsfield Unit (HU) and 61.4 HU for NR and UHR, respectively. The image noise levels were relatively constant across the range of CTDIvol values utilized in the patient images as shown in Figure 7b – consistent with previous phantom imaging results using the same reconstruction algorithm and kernel.[20]

Figure 7: Image noise results represented as (a) box plot comparisons between normal resolution and ultra-high-resolution “UHR” and (b) plotted separately as a function of CTDIvol. The red line in the middle of each box in (a) is the sample median, the blue bottom and top of each box are the 25th and 75th percentiles of the sample, and the red “+” signs are outliers. Statistically significant differences (P < 0.05) are indicated by an asterisk (*). (NR: Normal resolution, UHR: Ultra-high-resolution, CTDI: Computed tomography dose index.)

Export to PPT

DISCUSSION

The primary purpose of this work was to evaluate the efficacy of UHRCT for the characterization of lung nodules in comparison to conventional-resolution CT. We found subjective improvements in the characterization of lung nodules, supported by the findings of significantly higher margin clarity, IQ, and density confidence scores for UHRCT. Variability in quantification of the short and long axis of each nodule was not significantly different for UHRCT, but intra-reader agreement was moderate or good for these measurements. As anticipated, the increased spatial resolution in UHRCT led to significantly higher image noise, but it also improved the characterization of nodules in this study.

As with all newly introduced technology into clinical practice, it is important to assess the efficacy of such technology in comparison to existing, standard-of-care approaches. The difficulty in such comparisons is that they are usually limited to only a few sample patients, phantoms, or cadaver specimens. In some cases, the new technology is not directly compared to existing technology, since scanning a patient twice is likely not supported by each institution’s IRB. For example, Tsubamoto et al.[19] and Yanagawa et al.[15] explored the utility of UHRCT for the evaluation of pulmonary nodules in patients with lung disease, but they only scanned the patients using UHR acquisitions and did not compare directly against conventional CT. The same group also imaged cadaveric lungs, but images were only acquired on the UHRCT system in high-resolution modes.[8] This group also investigated the quality of lung imaging and compared directly between UHRCT and conventional CT, but they only imaged phantoms and cadaveric human lungs.[24] To address these limitations with previous studies, we utilized a validated simulation algorithm for generating NR reconstructions (NRsim). This methodology facilitated the comparison of two different technologies on the exact same patient. It is anticipated that this approach would be useful for other studies related to header, chest, and abdominal CT examinations.

More recently, research has been focused on investigating the utility of PCCT for lung imaging.[16,25] One group found improvements in the measurement accuracy of nodule volume and shape characterization using phantom scans with PCCT.[16] They then performed a prospective clinical study comparing PCCT and conventional CT for lung structure visualization. PCCT allowed radiologists to see higher-order bronchi and bronchial walls without compromising nodule evaluation.[25] While our research objectives differ from theirs, it is important to highlight a key difference in their reader study approach. Bartlett et al.[25] used a side-by-side reading design, presenting clinical reference images (i.e., conventional CT) and PCCT images (or UHRCT[15,19]) simultaneously, similar to other reader studies mentioned above.[15,19] In comparison, our study randomly displayed individual image datasets separately. We consider our approach more rigorous for assessing new imaging technology since the side-by-side method may introduce bias in the evaluation of nodule parameters if the readers become aware of which images are UHRCT (for example). In a realistic clinical setting, a radiologist will only have one technology available; therefore, we also believe the independent reading paradigm is more clinically realistic.

This study does have some limitations. First, the sample size was relatively limited at 104 lung nodules. All possible cases between February and November 2022 were included in this study, and as such the relatively limited sample size is a fundamental limitation of the frequency of non-contrast chest CT examinations performed at our institution on this scanner. Second, the training session used to train readers in using the visualization software, scoring the cases, and measuring the nodule size consisted of only three cases and was not necessarily selected to represent the large differences in margin clarity, IQ, nodule density, and nodule sized found in the cohort of cases includes in this work. This is the main reason we believe the intra-reader agreement was poor for margin clarity, IQ, and density confidence scores. In comparison, intra-reader agreement was moderate to good for size measurements as nodule measurement is a common task radiologists employ clinically. We believe that a more extensive training session would have been more effective in establishing consistency in individual readers. Ground truth characterization of the lung nodules based on pathologic evaluation was not obtained. Further investigation employing prospective patient recruitment and pathology may be beneficial for further exploration of the efficacy of UHRCT for lung cancer diagnosis. Furthermore, our study evaluated how, compared with conventional-resolution CT, ultra-high-resolution CT may influence the visualization of lung nodule imaging features by radiologists that are associated with or predictive of the likelihood of malignancy. However, future studies can be conducted to assess how this might potentially impact further management, such as interval follow-up CT imaging, positron emission tomography/CT, and tissue sampling recommendations.

CONCLUSION

UHRCT offers significant enhancements in lung nodules characterization in comparison to conventional-resolution CT. Ultra-high-resolution CT provides higher margin clarity, IQ, and density confidence scores in comparison to conventional-resolution CT. Variability in quantifying lung nodule short and long axes is not significantly different for ultra-high-resolution CT in comparison to conventional-resolution CT. Image noise is significantly higher for ultra-high-resolution CT in comparison to conventional-resolution CT, which can be attributed to the distinctly higher spatial resolution.