DOI: https://doi.org/10.21203/rs.3.rs-2322616/v1
Persistent subsolid nodules requiring follow-up are often detected during lung cancer screening; however, changes in their invasiveness can be overlooked owing to slow growth. The purpose of this exploratory study was to develop a method to automatically identify invasive tumors during multiple health check-ups.
We retrospectively reviewed patients who underwent screening using low-dose computed tomography (CT) between 2014 and 2019. Patients with lung adenocarcinomas manifesting as subsolid nodules resected after 5 years of follow-up were enrolled. The resected tumors were categorized into invasive or less-invasive groups. The annual growth or change rate (Δ) of the nodule voxel histogram on three-dimensional CT (e.g., tumor volume [cm3], solid volume percentage [%], mean CT value [HU], variance, kurtosis, skewness, and entropy) was assessed using radiomics. Multivariate regression modeling was employed to design a discriminant model.
Forty-seven tumors (282 detectable lesions over 5 years) were included (23 and 24 in the invasive and less-invasive groups, respectively). The median tumor volumes at the initial screening were 130 and 106 mm3 in the less-invasive and invasive groups, respectively; the difference was not significant (P = 0.489). In the multivariate regression analysis to identify the invasive group, Δskewness was an independent predictor (adjusted odds ratio, 0.021; P = 0.043). When combined with Δvariance (odds ratio, 1.630; P = 0.037), the assessment method had high accuracy for detecting invasive lesions (true-positive rate, 88%; false-positive rate, 80%).
During check-ups, close investigation by surgery for subsolid nodules can be suggested with the application of radiomics, particularly focusing on skewness.
Not applicable.
Early-stage lung cancer is commonly detected during lung cancer screening using low-dose computed tomography (LD-CT). Nevertheless, lung cancer remains the leading cause of death due to cancer worldwide [1]. Regarding lung cancer screening in the USA, the Lung Imaging Reporting and Data System (Lung-RADS) categories of tumor size and volume and the size of the solid component are widely used for management decisions [2]. Furthermore, LD-CT is commonly used in Europe to assess tumor volume [3]. Despite standardizations for management decisions based on tumor size and volume [2, 3], slow-growing tumors, which tend to be overlooked during screening, become more invasive than expected. For patients with a positive result on screening with LD-CT, high-resolution CT and/or transbronchial lung biopsies are generally performed. Additionally, a close investigation by surgery was performed when small-sized subsolid nodules were present in the periphery of the lungs. However, non-specialists may struggle with providing a referral to a thoracic surgeon with the recommended screening methods for slow-growing tumors [2, 4]. In contrast, our previously reported radiologic technology could have the potential to screen for changes in the invasiveness of lung adenocarcinomas and automatically notify about an increased need for surgery [5]. Thus, in this exploratory study, we focused on slow-growing tumors. Through radiomics based on the voxel histogram of serial LD-CT, we aimed to develop a new method that identifies tumors with increasing invasiveness that require surgery among subsolid nodules during health check-ups.
This single-center observational retrospective study was approved by the Institutional Review Board of St. Luke’s International Hospital, Tokyo, Japan (approval no. 19-R026; issued April 9, 2020). The need for written informed consent was waived. A total of 201,275 patients underwent periodic health examinations between January 2014 and 2019. After 51,136 patients were examined using chest CT, 309 were suspected of having malignancies and were referred to the outpatient clinic for close surgical investigation. Among these patients, 115 surgeries were performed for nodules suspected to be malignant. We excluded patients with (i) no available data on annual LD-CT in the last 5 years (n = 53) to minimize measurement bias and (ii) a pathological diagnosis, except for lung adenocarcinoma (n = 15). Finally, 47 patients (47 × 6 = 282 subsolid nodules) were enrolled (Fig. 1). The study was conducted in accordance with the Declaration of Helsinki.
LD-CT was performed using a 64-detector row scanner (Aquilion ONE, Toshiba Medical Systems, Tokyo, Japan; Revolution and Optima 660, GE Healthcare Japan, Tokyo, Japan) with the following standard parameters: 120 kV, automatically set for amplification, and bone reconstruction algorithm. A 2.5-mm slice thickness was acquired for all images using standard reconstruction kernels with lung window settings (window level, − 500 Hounsfield units [HU]; window width, 1500 HU). The acquired images were evaluated by expert radiologists (DY with 5 years of experience, MM and YK with > 20 years of experience each).
Data of 282 pulmonary lesions (47 nodules per year) detected on LD-CT were extracted semi-automatically. The volume of interest (VOI) was estimated using a three-dimensional image analysis software (SYNAPSE VINCENT; Fujifilm Medical, Tokyo, Japan). This was equivalent to the overall tumor volume determined using voxels (mm3) and CT values (HU). We set the ratio of the solid component (within − 300 HU) expressed as the volume percentage (% solid), and the area of ground-glass opacity (ranging from − 1000 HU to − 300 HU) had a border of − 300 HU (Fig. 2). Subsequently, a voxel-based histogram analysis (VHA) was conducted for the VOI by calculating the following five parameters, as previously described [5]: mean CT value, variance, skewness, kurtosis, and entropy. We focused on changes in the tumor from the initial detection until the most recent check-up after 5 years.
Therefore, the growth rate or change rate (Δ) of seven radiological parameters (X) based on VHA was calculated and evaluated using the following formula:
ΔX per year = value at the time of surgery - mean value for four years from the initial detection to one year prior to surgery (mean of five times)
X = tumor volume (cm3), solid volume percentage (% solid, %), mean CT value (HU), variance (×104), kurtosis, skewness, or entropy.
The 8th edition of the Tumor, Node, Metastasis staging system was applied for this study [6]. According to the International Association for the Study of Lung Cancer, American Thoracic Society, and European Respiratory Society classification of lung adenocarcinoma, 47 resected specimens were classified into the following two groups: the less-invasive group (adenocarcinoma in situ [AIS] and minimally invasive adenocarcinoma [MIA]) and invasive group (invasive adenocarcinoma [IA]) [6]. The measurements of the tumor dimensions (maximum and solid sizes in diameter) for all 282 lesions performed by expert radiologists were applied as a reference to assess the clinical T factors. The consolidation-to-tumor ratio (CTR) was also determined based on the findings.
Segmentectomy was performed under the following conditions: i) nodules < 2 cm in size, ii) pure ground-glass (CTR = 0) or part-solid nodules on CT (CTR < 0.5), and iii) suspected metastases from other organ cancers [7].
We calculated descriptive statistics for patient characteristics and continuous data in terms of the growth rate and change rate (Δ) of the seven radiological parameters. Categorical variables were summarized as numbers and proportions and continuous variables as medians and interquartile ranges (IQR). To compare each continuous variable between the less-invasive and invasive groups, we performed a univariate analysis based on the Mann–Whitney U test. All data were analyzed using two-sided hypothesis tests. Statistical significance was set at p < 0.05. To provide a visual analysis of how each parameter changed over time, we plotted all patient data per year for each parameter and applied locally weighted scatterplot smoothing (LOWESS) to capture the features of the changing trends. We proposed a new approach in terms of the prediction of increasing invasiveness based on Δ. Therefore, to validate the results derived from our new method, we performed a visualization analysis using LOWESS.
We also performed multiple logistic regression analysis (adjusted) and identified the essential factors needed to detect the invasive group by exploiting the differences (Δ) in each of the seven variables between the two groups. We performed stepwise variable selection for the logistic regression model based on the seven variables. Finally, we estimated the most sophisticated logistic regression model after variable selection.
Using the final logistic regression model, we calculated the area under the receiver operating characteristic curve (AUC) to determine the threshold score. A decision tree of the classification and regression tree (CART) model was constructed to predict the IAs and estimate the cut-off values of the critical variables. To conduct the CART analysis, we used the Gini index criterion and set the tuning parameter for the decision-tree complexity to 0.01. All descriptive statistical analyses, univariate analysis, LOWESS, multiple logistic regression, and CART analyses were conducted using R (version 3.4.3; R Foundation for Statistical Computing, Vienna, Austria).
Forty-seven consecutive patients were reviewed (one pulmonary lesion per patient). Following the medical check-up, the patients were referred for possible malignancy and surgery with diagnostic and curative intent. The possibility of lung cancer was unknown 5 years before surgery. The median age of the patients was 68 years [IQR 64–73.5], with a standard body type. The male to female ratio was approximately 1:1 (23 males, 48.9%). A few patients were diagnosed with comorbidities such as chronic obstructive pulmonary disease (n = 2, 4.2%) and interstitial pneumonia (n = 1, 2.1%). Approximately 43% (n = 20/47) of patients were smokers.
Table 1 shows the clinicopathological features of resected subsolid nodules. All 47 pulmonary lesions were excised using thoracoscopy, with sufficient margins measuring > 2 cm. The subsolid nodules displayed slow growth over the long term; therefore, the majority of patients underwent sublobar resections (n = 36/47, 76.6%). During surgery, no significant difference was observed in the maximum tumor size between the less-invasive (12.7 [IQR 9.0–15.0] mm) and invasive (13.0 mm [IQR 9.2–16.2] mm; P = 0.549) groups. All the lesions corresponded to lung adenocarcinomas at clinical stages 0 and IA. Pathological diagnoses displayed stages N0, M0, 0, and IA in all the cases. Among the 47 resected nodules, 14 (29.8%) were classified as AIS, and 10 (21.3%) were classified as MIA. The remaining 23 (48.9%) resected nodules were classified as IAs.
| Clinicopathological and operative data (n = 47) | |
|---|---|
| Localization of tumors | |
| Right/left | 29/18 (61.7/31.3%) |
| Upper/middle/lower | 24/4/19 (51.1/8.5/40.4%) |
| Diameter of tumors on preoperative CT | |
| Maximum size, mm | 13.7 (9.0–16.1) |
| Solid size, mm | 6.4 (0.0–8.9) |
| Consolidation-to-tumor ratio | 0.3 (0.0–0.8) |
| Clinical T factor (N0M0) | |
| Tis/T1mi | 19/8 (40.4/17.0%) |
| T1a/T1b/T1c | 11/6/3 (23.4/12.8/6.4%) |
| Pathological T factor (N0M0) | |
| Tis/T1mi | 14/10 (29.8/21.3%) |
| T1a/T1b/T1c | 16/5/2 (34.0/10.6/4.3%) |
| Histological type | |
| AIS | 14 (29.8%) |
| MIA | 10 (21.3%) |
| Lepidic-predominant | 3 (6.4%) |
| Acinar-predominant | 3 (6.4%) |
| Papillary-predominant | 15 (31.9%) |
| Others (solid or invasive mucinous) | 2 (4.3%) |
| Surgical procedures | |
| Wedge resection | 8 (17.0%) |
| Segmentectomy | 28 (59.6%) |
| Lobectomy | 11 (23.4%) |
| Abbreviations: AIS, adenocarcinoma in situ; CT, computed tomography; MIA, minimally invasive adenocarcinoma. | |
| Data are reported as the median (interquartile range) or count (percentage). | |
A typical histogram transition for a subsolid nodule located in the right lower lobe is shown in Fig. 3. The gray scales on the CT images represent tumor conversion to radiological information on voxel histograms based on three-dimensional images. Applying the methodology of a previous study [5], 282 lesions were assessed under seven radiological parameters and followed up for 5 years; these were calculated and illustrated in Fig. 4. The median values of the initial maximum diameter and tumor volume were 5.8 (IQR 4.0–9.1) mm and 130 (IQR 81.5–256.8) mm3, respectively, in the less-invasive group and 4.2 (IQR 3.2–8.2) mm and 106 (IQR 59.1–211.5) mm3, respectively, in the invasive group. No significant differences were observed in the maximum tumor size (P = 0.139) and volume (P = 0.489) at the initial screening. The growth rate of 45 lung tumors < 523 mm3 (i.e., 10 mm in diameter according to the classification of the Lung-RADS score) at the initial screening is shown in Fig. 5. No association was observed between initial volume and growth rate. Furthermore, Table 2 summarizes the variations in the parameters in the less-invasive and invasive groups. All radiological factors, except tumor volume and mean CT value, showed significant differences between the two groups.
| Characteristics | Less-invasive group (n = 24) | Invasive group (n = 23) | P-value |
|---|---|---|---|
| Δ tumor volume, cm3 per year | 0.406 (0.113, 1.046) | 0.522 (0.236, 1.017) | 0.416 |
| Δ % solid, % per year | 1.590 (0.000, 4.360) | 9.620 (1.060, 20.690) | 0.004 |
| Δ mean CT value, HU per year | −12.530 (− 31.337, 40.019) | 33.379 (− 19.474, 110.517) | 0.099 |
| Δ variance, ×104 per year | 0.639 (− 0.003, 1.187) | 1.435 (0.735, 3.403) | 0.016 |
| Δ kurtosis, per year | 0.536 (− 0.108, 1.944) | −0.228 (− 0.934, 0.319) | 0.007 |
| Δ skewness, per year | 0.577 (0.243, 0.838) | −0.073 (− 0.420, 0.350) | < 0.001 |
| Δ entropy, per year | 0.499 (0.276, 0.688) | 0.924 (0.500, 1.271) | 0.018 |
| Abbreviations: CT, computed tomography; Δ, delta (amount of change); HU, Hounsfield units; % solid, solid volume percentage; 3D, three-dimensional. | |||
| Data were analyzed using the Mann–Whitney U test and are reported as medians (interquartile ranges). | |||
The factors related to the detection of malignant transformation during multiple screenings using univariate and multivariate analyses are summarized in Table 3. The change in skewness was the most critical independent screening parameter for IAs, following adjustment for confounders. There was no multicollinearity among the seven factors. In the multivariate analysis, Δskewness (odds ratio [OR] 0.096, 95% confidence interval [CI] 0.020–0.450; P = 0.003) and Δvariance (OR 1.630, 95% CI 1.030–2.58; P = 0.037) remained significant detectable factors in the invasive group. The AUC was 0.84 (95% CI 0.721–0.960], indicating a sensitivity of 73.9%, a specificity of 87.5%, and an accuracy of 80.9%. Utilizing a CART model to construct a decision tree, the variation in asymmetry was detected to constitute the first step to screen out, followed by variance. The cut-off values for Δskewness and Δvariation per year were − 0.21 and 1.09, respectively. If a subsolid nodule measured − 0.21 or less for Δskewness and less than 1.09 for Δvariance per year, the tumor was positively associated with increased invasiveness during extended follow-ups.
| Univariate logistic regression model | Multivariate logistic regression model | |||||
|---|---|---|---|---|---|---|
| (Crude) | (Adjusted) | (After variable selection) | ||||
| Characteristics | OR (95% CI) | P-value | OR (95% CI) | P-value | OR (95% CI) | P-value |
| Δ tumor volume, cm3 per year | 1.360 | 0.312 | 0.877 | 0.742 | ||
| (0.750–2.460) | (0.402–1.910) | |||||
| Δ % solid, % per year | 1.130 | 0.011 | 1.040 | 0.737 | ||
| (1.030–1.250) | (0.814–1.340) | |||||
| Δ mean CT value, HU per year | 1.010 | 0.076 | 0.998 | 0.819 | ||
| (0.999–1.010) | (0.978–1.020) | |||||
| Δ variance, ×104 per year | 1.530 | 0.026 | 1.440 | 0.372 | 1.630 | 0.037 |
| (1.050–2.230) | (0.647–3.200) | (1.030–2.580) | ||||
| Δ kurtosis, per year | 0.544 | 0.028 | 1.900 | 0.316 | ||
| (0.317–0.936) | (0.541–6.690) | |||||
| Δ skewness, per year | 0.107 | 0.002 | 0.021 | 0.043 | 0.096 | 0.003 |
| (0.026–0.445) | (0.001–0.889) | (0.020–0.450) | ||||
| Δ entropy, per year | 4.190 | 0.025 | 2.45 | 0.388 | ||
| (1.190–14.700) | (0.320–18.800) | |||||
| Abbreviations: CI, confidence interval; CT, computed tomography; HU, Hounsfield units; OR, odds ratio; % solid, solid volume percentage | ||||||
| 21 | ||||||
| Abbreviation: CT, computed tomography | ||||||
| Abbreviations: CT, computed tomography; HU, Hounsfield units | ||||||
| Abbreviations: CT, computed tomography; HU, Hounsfield units; LD-CT, low-dose computed tomography. | ||||||
| Abbreviations: CT, computed tomography; HU, Hounsfield units; LD-CT, low-dose computed tomography. | ||||||
| Abbreviations: CT, computed tomography; Lung-RADS, Lung CT Screening Reporting and Data System | ||||||
In this exploratory study on subsolid nodules of early-stage lung adenocarcinomas, we propose a new method for detecting invasive lesions that should be closely investigated by surgery during long-term follow-up. VHA was performed for serial LD-CT images of pathological lung adenocarcinomas. Consequently, the variation in skewness over time was more significant for subsolid nodules than for radiological factors such as volume or solid percentage. Thus, our proposed method has the potential to better clarify changes in invasiveness.
Although the growth process of early-stage lung adenocarcinoma is not fully elucidated, tumor size and solid components are known to increase with growth from atypical adenomatous hyperplasia to IAs [6]. To date, the management of lesions detected by lung cancer screening has been recommended based on tumor size and volume and the size of the solid component, as per the Lung-RADS and National Comprehensive Cancer Network guidelines [2, 4]. However, we occasionally observed lesions with uniformly increased internal density suspected of having increased invasiveness without changes in the tumor size or solid components on LD-CT. Therefore, we hypothesized that radiological factors besides tumor size and solid components are necessary to assess invasiveness during the follow-up of subsolid nodules. We decided to focus on subsolid nodules over time to enable a comprehensive analysis of various radiological features.
The study population included 47 patients who were followed-up for 5 years. Despite the limited number of patients, these findings allowed us to understand the growth of lung adenocarcinomas manifesting as subsolid nodules. In our health screening facility, the detection rate of nodules on LD-CT was 0.6% (n = 309/51,136). Of the 309 patients/nodules, 115 underwent surgery. Nearly all pathological diagnoses were lung adenocarcinomas (87%; n = 100/115), and approximately half (41 subsolid nodules and 12 solid nodules) were resected after 1 year of follow-up. Traditionally, a few slow-growing tumors have required long-term follow-up through radiographic screening. However, slow-growing tumors are frequently detected during health check-ups with LD-CT, and physicians at annual health check-ups appear to find it difficult to determine when to consult surgeons regarding subsolid nodules. Our new method provides the first step toward resolving this issue.
Recently, there has been growing interest in radiomics studies. Substantial volumes of medical imaging data have been comprehensively analyzed, and the relationship between medical imaging data and clinical information has been investigated [8–10]. In lung cancer research, radiomics has been utilized for diagnosis and the prediction of prognosis and response to treatment [11, 12]. Previously, we reported radiomics based on VHA with a single CT scan taken before surgery for lung adenocarcinoma, which could accurately assess malignancy [5]. Among the seven radiological factors in this study, those that assessed the level of invasiveness were, in ascending order, tumor volume, % solid, kurtosis, and entropy. Subtle differences in images invisible to the naked eye were clarified in a previous study [5]. Other investigators have also demonstrated the usefulness of VHA for lung lesions [13, 14]. In this study, we used the methodology of our previous study to investigate the differences between less-invasive and invasive lung adenocarcinomas manifesting as subsolid nodules. The seven factors utilized in our previous study were followed up over time to assess subsolid nodules with slow growth. These were detected during annual check-ups by using serial LD-CT. Changes in tumor volume and solid components (% solid) were not sufficient to evaluate the change in invasiveness among subsolid nodules; however, with a focus on skewness in the voxel histogram, changes in invasiveness became clearer. In particular, skewness declined gradually in the invasive group but increased in the less-invasive group. Such a decrease in skewness implies a shift to the right in the histogram; that is, the decrease in skewness reflects an increase in internal density, displaying invasiveness in subsolid nodules and a slight change in tumor volume. Thus, for subsolid nodules with slow growth that do not meet the criteria for changes in tumor size or volume or in the size of the solid component during an annual check-up with LD-CT, the newly proposed evaluation focusing on skewness on VHA will prevent overlooking invasiveness and allow a timely referral for a close investigation by surgery. Conversely, watchful waiting without surgery may be recommended for suspected lesions in patients with less-IAs diagnosed using this method.
There are some limitations in the study. First, this was a retrospective, single-center, exploratory study. To validate the results of this study, a validation set with a different patient cohort would be necessary. Second, a limited number of tumors were included because lesions displaying significant changes within a few years were excluded to minimize the measurement bias. However, owing to the rarity of slow-growing adenocarcinomas resected after 5 years of follow-up, the data obtained were of considerable importance. Third, the study object was exclusively pathological lung adenocarcinoma. In lung cancer screening with LD-CT, the utility of the study method in differentiating between benign and malignant lesions needs to be assessed. Efforts should be made to reduce the number of false positives for benign lesions during screening, and our method with serial LD-CT may lead to a more accurate evaluation; however, this was beyond the scope of this study.
In lung cancer screening with LD-CT, close investigation by surgery can be recommended for subsolid nodules followed up over time with the application of radiomics, particularly focusing on skewness.
AIS: adenocarcinoma in situ; AUC: area under the receiver operating characteristic curve; CART: Classification and Regression Trees; CT: computed tomography; CTR: consolidation to tumor ratio; HU: Hounsfield unit; IA: invasive adenocarcinoma; IQR: interquartile range; LD-CT: low-dose computed tomography; LOWESS: locally weighted scatterplot smoothing; Lung-RADS: Lung Imaging Reporting and Data System; MIA: minimally invasive adenocarcinoma; OR: odds ratio; VHA: voxel-based histogram analysis; VOI: volume of interest
Ethics approval and consent to participate
The study design was approved by the Institutional Review Board of St. Luke’s International Hospital, Tokyo, Japan (approval no. 19-R026; issued April 9, 2020). The need for written informed consent was waived by the Institutional Review Board of St. Luke’s International Hospital owing to the retrospective study design. All the methods were followed in relevant guidelines and regulations.
Consent for publication: Not applicable.
Availability of data and materials: The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request.
Competing interests: None of the authors has any conflicts of interest.
Funding: None.
Authors’ contributions
NY was a major contributor in writing the manuscript. NY and DY collected the data. NY and KH analyzed the data. FK and TB mainly revised the article. TB supervised the design of this paper. All authors read and approved the final manuscript.
Study conception: NY, FK, TB
Data collection: NY, DY
Analysis: NY, KH
Investigation: NY, DY, KH
Manuscript preparation: NY
Critical review and revision: all authors
Final approval of the article: all authors
Acknowledgments: The authors thank Dr. Yasuyuki Kurihara, Dr. Masaki Matsusako, and all the staff who contributed to this study. We would like to thank Editage [http://www.editage.com] for editing and reviewing this manuscript for English language use.
Authors’ information (optional): Not applicable.