1. Introduction
Lung cancer is one of the deadliest cancers worldwide. It accounts for 18% of the cancer deaths overall and is a leading cause of cancer death in men and the second leading cause in women [
1].
CT screening examinations enable the detection of small lung cancers in the early stages, improving the patient outcome substantially [
2,
3]. After detection of the nodules, core-needle biopsy can be used to determine lung cancer types and subtypes, a relatively safe procedure facilitating optimal targeted therapy for each patient [
4,
5].
Among others, the National Lung Cancer Screening Trial (NLST) found a reduction in mortality rates by 20% in a high-risk population (n = 53,454) when using low-dose CT (LDCT) instead of chest radiography for lung cancer screening (LCS) [
6]. Previous investigations also showed that chest radiography is significantly inferior to chest CT examinations regarding the detection of small pulmonary tumors [
7,
8]. The importance of detecting lung cancers in the tumor stage 1 (T1) is reflected by the 5-year survival rate, which drops from 61% in localized tumor stages to 7% in advanced tumor stages [
9]. Therefore, the current study focused on T1 tumors, which by definition are not larger than 3 cm across, have not reached the membranes surrounding the lungs, and do not affect the main branches of the bronchi.
Still, an immanent issue with LCS is the exposure to ionizing radiation. Therefore, the optimization of screening CT protocols plays a key role in the field and has driven numerous investigations in the past. In particular, in the past decade, with the introduction of iterative reconstruction algorithms [
10] and highly efficient detector assemblies, LDCT has become a reality that is already in broad clinical use, especially in the field of LCS, with the UK currently leading the way [
11,
12,
13,
14].
In the context of LCS, the minimization of radiation exposure is particularly relevant since a large number of healthy individuals without symptoms are exposed. Several studies have addressed this issue in the past using different approaches, and there is a broad consensus that dose reduction is feasible [
15,
16,
17,
18].
In this context, LDCT protocols are commonly associated with effective doses of one to two mSv at best [
19]. However, recent investigations have shown that further dose reduction is feasible, with an effective radiation dose well below one mSv. Our institution has built a profound expertise regarding ULDCT examinations in phantom studies, which showed that low-dose (1–2 mSv) and ultralow-dose (0.1–0.2 mSv) [
20] examinations are feasible for detecting small solid and subsolid nodules [
17], with rather high sensitivities and specificities compared to the original dose examinations [
21].
Another noteworthy new development in this context is the photon-counting detector CT, as it offers the opportunity to perform chest CT scans at very low doses with minimal noise, retaining a comparable image quality. The first studies regarding lung nodule detection and classification reported very promising results in humans and phantoms [
22,
23], but further research is warranted, e.g., to rule out potential impacts on patient management recommendations and to elaborate on optimized protocols for indications, such as LCS. However, it underlines the importance of exploring the potential effects of CT dose reduction on pulmonary nodule management.
Several studies have evaluated the effect of dose reduction on the detection of pulmonary nodules in well characterized but relatively small clinical cohorts [
24,
25], and the validation of those results over a wide variety of different vendors and scanners in a larger clinical cohort is still missing.
This study aimed to evaluate the feasibility of ultra-low dose protocols regarding the detection and classification of histologically proven pulmonary T1 cancers. Unlike previous studies, it utilizes a highly heterogeneous cohort, including data from different sites, including various scan protocols and CT scanner types. Furthermore, it aimed at assessing the potential impact on patient management caused by dose reductions by comparing shifts in the hypothetical Lung CT Screening and Reporting System (LungRADS) scores between the different dose-level groups.
2. Materials and Methods
This is a retrospective study with a fully crossed block design with multiple readers and multiple cases. It was approved by the local ethics committee and conducted in accordance with the principles of the Declaration of Helsinki. Only patients with written informed consent were included in the cohort provided by the local lung cancer center (LCC). The authors had full access to and take full responsibility for the integrity of the data.
2.1. Study Cohort
The study cohort was based on a repository provided by the local lung cancer center (LCC). It consisted of 218 individuals, all with histologically proven T1 tumors of the lung and available chest CT scans. The examinations were synchronized with the institutional picture archiving and communication system (PACS) and collected in a private user case list. The relevant dose parameters, such as dose-length-product (DLP) or the CT dose index volume (CTDIvol), were documented for each CT scan. Inclusion criteria were as follows: resected pulmonary lesion, histopathologic diagnosis of lung cancer, lesion size <3 cm (T1-stage), preoperative CT-scan present, patient age >18 years. Exclusion criteria were as follows: absence of preoperative CT-scan, higher tumor stages, explicit denial of further data use, insufficient image quality.
The examinations of six patients were deemed ineligible for virtual dose reduction due to excessive image noise (n = 4) and missing CT slices (n = 2). Additionally, two more scans were excluded because of incomplete lung coverage, to avoid potential bias from the smaller scan volume. The characteristics of the included patients (n = 210) and tumors (n = 250) are shown in
Table 1.
2.2. CT Acquisition and Creation of Virtual (Ultra)Low-Dose Protocols
The 218 CT examinations originated from over 20 different sites all over the country with 5 different CT manufacturers (Siemens, n = 130; Philips, n = 34; GE, n = 28; Toshiba, n = 25; Canon, n = 1). These examinations were conducted over an 8-year period (2010–2018), with the DLP and CTDI values remaining within the respective National Diagnostic Reference Levels [
26]. The acquired minimum slice thickness varied from 0.5–3 mm with the vast majority (>80%) being ≤1.5 mm. The reconstruction algorithms contained filtered-back projection (n = 130, 60%), as well as iterative reconstruction (IR, n = 88, 40%). Scan volumes of the included examinations contained whole-body examinations (n = 4, 2%), Positron Emission Tomography and Computed Tomography (PET/CT) scans (n = 36, 17%), chest plus abdomen or neck (n = 46, 22%), and chest-only acquisitions (n = 124, 59%) (
Table 2).
Regarding the scan protocol, 62% of the examinations were performed on the local 128-row multidetector Flash CT scanner (Siemens SOMATOM Definition Flash, Siemens Healthineers, Erlangen, Germany) featuring iterative reconstruction algorithms (iterative reconstruction in imaging space (IRIS)) and a detector system with integrated readout electronics, a gantry rotation time of 0.28 s, and a pitch of 2.2. For image reconstruction, an I30f soft tissue kernel was utilized; iterative reconstruction was used to create axial stacks of a 1 mm slice thickness.
The anonymized examinations were transferred to a dedicated post-processing imaging lab specialized in LDCT simulations. The reduced dose simulations were produced by adding statistical noise to the images using a previously described CT image-based noise addition tool [
27]. Four reduced dose level simulations were created out of every CT scan, leading to five different dose levels for each examination: 100% (original), 50%, 25%, 10%, and 5% doses.
These simulations were consecutively re-transferred to our institution.
Table 2.
CT parameters.
Dose parameters of the original scans, mean (SD) | |
DLP, mGycm | |
Chest only (n = 124) | 315.6 (213.5) |
Chest plus neck/abdomen (n = 46) | 892.6 (557.4) |
Whole-body acquisition (n = 4) | 540.8 (501.2) |
PET-CT (n = 36) | 308.6 (155.6) |
CTDIvol, mGy | |
Chest only (n = 124) | 13.7 (13.4) |
Chest plus neck/abdomen (n = 46) | 27.2 (23.9) |
Whole-body acquisition (n = 4) | 16.4 (25.1) |
PET-CT (n = 36) | 4.1 (2.1) |
Effective dose, mSv # | |
Chest only (n = 124) | 4.4 (3.0) |
Chest plus neck/abdomen (n = 46) | 12.5 (7.8) |
Whole-body acquisition (n = 4) | 7.6 (7.0) |
PET-CT (n = 36) | 4.3 (2.2) |
Slice thickness, n (%) | |
0.5 mm | 4 (1.2%) |
0.625 mm | 3 (0.9%) |
0.75 mm | 1 (0.3%) |
0.9 mm | 9 (2.6%) |
1 mm | 85 (25.0%) |
1.25 mm | 29 (8.5%) |
1.5 mm | 12 (3.5%) |
2 mm | 48 (14.1%) |
2.5 mm | 2 (0.6%) |
3 mm | 16 (4.7%) |
4 mm | 1 (0.3%) |
2.3. Pilot Study
After preparation of the reduced dose simulations, each reader would have needed to review 1050 examinations (210 × 5), which would have resulted in a highly time-consuming task. Since the differences between adjacent dose levels visually did not seem very striking, a pilot study was conducted in order to compare the signal-to-noise-ratios (SNRs) and contrast-to-noise-ratios (CNRs) of the five dose level groups (study workflow depicted in
Figure 1).
Therefore, a board-certified radiologist (AAP) measured the attenuation in Hounsfield units (HUs) and the corresponding standard deviations (SDs) of the air outside the patient (anterior of the sternum), in the bone (central in the vertebral body), and in the soft tissue (autochthonous back musculature) above the level of the diaphragm in 10 randomly chosen patients. SNR was calculated by dividing the signal intensity (SI) of the soft tissue by the background noise (SD of soft tissue SI), CNR by dividing the difference between the SI of bone, and soft tissue by the soft tissue background noise as follows:
2.4. Readout
Two independent blinded readers (reader 1 with seven years and reader 2 with five years of experience in chest radiology) conducted the readouts of the main study on dedicated workstation monitors (BARCO Coronis Fusion 6MP LED, Kortrijk, Belgium). The readers scored all eligible examinations (3 × 210 = 630 examinations) in a randomized order regardless of the dose level, rating the nodule diameter (based on the categories following LungRADS v2022) and the location and density category (solid, part-solid, ground-glass) of each nodule on a spreadsheet. The readers were allowed to use all kinds of tools, such as multiplanar reformats or maximum intensity projections, in order to read the examinations in the most realistic setting possible. A board-certified radiologist with seven years of experience in chest radiology read all examinations independently and documented the presence of other pulmonary diagnoses, such as emphysema, fibrosis, or pneumonia.
2.5. Statistical Analysis
Metric variables are reported as the mean (standard deviation), with categorical variables as absolute numbers (relative percentage). In the pilot study, the SNR and CNR of the different groups were compared using the Friedman test for multiple not normally distributed paired samples with Bonferroni correction for multiple comparisons. For the main study, a generalized linear mixed model (GLMM) with crossed random intercepts for readers and lesions and the dose reduction level as a fixed effect was designed. The binarized endpoints (yes/no) were correct nodule detection, correct categorization of nodule attenuation, size, and localization.
In a subgroup analysis, each reader’s results regarding the respective endpoints for each (simulated) dose level were compared using the Cochran’s Q test.
To assess the possible clinical impact of the findings, the hypothetical LungRADS scores of the tumors were calculated for every dose level. Shifts between the risk groups, which could be assumed to be caused solely by a dose reduction, were documented. The scores were based on LungRADS v2022.
Interrater agreement was assessed by using Cohen’s Kappa (κ). According to Landis and Koch, kappa-values of 0.00 to 0.20 were interpreted as slight, 0.21 to 0.40 as fair, 0.41 to 0.60 as moderate, and 0.61 to 0.80 were interpreted as substantial, while values between 0.81 and 1.00 were interpreted as almost perfect agreement [
29].
A p-value of <0.05 was considered statistically significant. All analyses were performed using dedicated software: SPSS (SPSS Statistics, IBM Corp., version 25.0. Armonk, NY, USA) and GraphPad Prism (GraphPad Software, Inc., version 8, San Diego, CA, USA).
4. Discussion
This study analyzed the impact of dose reduction on the detection, localization, categorization, and management of pulmonary T1 tumors by using virtual (ultra)low-dose CT protocols.
The main finding of this study was that dose reduction in chest CT is feasible regarding pulmonary nodule detection, localization, and classification. However, according to the subgroup analyses, tumor localization and size categorization might be affected by a dose reduction for certain readers. Fletcher et al. analyzed the detectability of pulmonary nodules in the chest CT scans of 21 patients containing 28 nodules using five different dose levels and found that dose reductions by 70% or more are non-inferior compared to the routine clinical dose levels [
16]. In a follow-up study, they revealed that scanning pulmonary nodules 5 mm or larger at very-low-dose levels are feasible down to a 10-quality reference mAs (QRM) level but might lead to a loss of detection regarding a significant proportion of part-solid nodules [
24].
In the current study, there were no differences regarding the detection rates of part-solid nodules between the different dose levels; however, it has to be mentioned that the number of included part-solid nodules was rather small (n = 28).
The mean CTDIvol in this study was relatively high compared to similar studies, most probably because the included examinations comprised not only chest CT scans but also abdominal or whole-body scans acquired in various clinical settings [
24]. However, the effectively applied doses were in similar ranges [
30].
Nodule detection rates of both readers were excellent for all dose levels while maintaining an acceptable FPR (range: 0.13–0.45).
In order to evaluate the secondary endpoint, readers had to localize the nodules correctly and categorize them by size and attenuation; the categories were defined in accordance to LungRADS v2022 in order to assess the potential impact on patient management. LungRADS categories shifted between the different doses, indicating a potential impact on patient management. After calculation of the hypothetical LungRADS scores, 15.2% and 16.8% of all tumors would have shifted to a different LungRADS category after dose reduction from the original dose to the 25%- and 5%-dose level, respectively, for reader 1. The corresponding values for reader 2 were even a bit higher with 18.0% and 20.0%. Taking into regard a measurement variability of 25% reported by several in vivo “coffee-break” studies and the proportion of differing LungRADS scores between two different computer-aided diagnosis (CAD) systems measuring the same nodule of approximately 15%, these values seem acceptable and are in line with the literature [
31,
32,
33].
In a similar study, Paks et al. compared an LDCT protocol to an ULDCT protocol regarding pulmonary nodule detection and volumetry in 188 solid pulmonary nodules greater than 2 mm and concluded that ULDCT delivers comparable results and therefore may be used for follow-up examinations [
30]. However, they did not assess the impact on patient management specifically, limiting the comparison to the current study.
Hata et al. and Milanese et al. both assessed the impact of dose reduction on LungRADS classification by radiologists [
34,
35]. While Milanese et al. found excellent intraobserver agreement between low dose and ultra-low dose scans, Hata et al. reported varying agreements between the original dose and reduced dose scans, indicating a potential impact of dose reduction on the LungRADS classification, which is in line with the current results. However, it should be noted that in the latter study, the median volume of the nodules was approximately 50 mm³, compared to volumes ranging from 75 mm³ to 194 mm³ in Milanese et al., which made interpretation more challenging.
A task for future studies will be the evaluation of (AI-) CAD systems in this context, since such tools are broadly utilized in daily clinical routines in order to support radiologists, who are confronted with an increasing workload and benefit from CAD systems, especially if used as second reader devices [
36,
37,
38]. Regarding the current study, it can be hypothesized that the use of (AI-) CAD systems potentially could have enhanced the readers’ performance, especially reader 1 might have had benefitted from a second reader device. In theory, the use of CAD systems leads to more robust and reproducible readout results.
Interestingly, it could be shown that CT dose reduction has an influence on the performance of a deep learning (DL)-based CAD system regarding malignancy prediction in a high-risk cohort of proven malignancies [
39]. This finding implies that the CAD systems still require radiological supervision as their performance is dependent on image quality.
For daily clinical routines, these results imply increased awareness while reading reduced-dose chest CT scans, for instance in the context of LCS. This applies to both readouts performed by radiologists alone and AI-assisted readouts. If CAD systems are utilized, the results demand supervision by a radiologist in a second reader scenario, since the risk of lung cancer under- or overestimation and all the associated, potentially unnecessary consequences, such as biopsies, is given.
This study has limitations. First, although the results of the current study indicated no statistical difference between the different dose level scans, the design did not allow for any statement on the non-inferiority of the reduced-dose CT scans, which should be targeted in follow-up studies.
Second, the virtual reduced-dose protocols were created during post-processing and are not perfectly transferable to true ULDCT protocols. However, this approach was the most realistic and at the same time ethically justifiable one. Third, CT scans from various institutions and vendors were included with differing settings, which may have affected the quality of the simulations. On the other hand, this approach led to robust and generalizable results. Fourth, there were only a small number of part-solid nodules included in this study. This type of nodule is very relevant, since it has a higher probability of malignancy and is specifically vulnerable to higher image noise. However, the management defining and therefore most relevant part of the nodule is the solid portion, which was evaluated by the readers. Fifth, the results were based on the readout results of only two readers resulting in limited generalizability and demanding follow-up studies with a higher number of readers. Lastly, there was no washout period for the readers between the readout of the different dose levels to definitely rule out recall bias. However, randomization of the large number of cases prevented an order effect.
In conclusion, the results of this study indicate that a dose reduction to 25% or 5% of the original dose is feasible for the detection and localization of pulmonary T1 cancers. Alterations to patient management based solely on dose reduction cannot be ruled out by the current results; however, there is no clear tendency towards malignancy risk over- or underestimation. For clinical routines, respective measures should be taken to address this problem, for instance the utilization of a second reader setup.