Background
Surgery with curative intent is the foundation for management of early-stage non-small-cell lung cancer (NSCLC). However, 10–20% of stage IA and 30% of stage IB patients will die within 5 years of surgery [
1]. For stage IA patients, adjuvant chemotherapy is not recommended by American Society of Clinical Oncology (ASCO) and National Comprehensive Cancer Network (NCCN) guidelines [
2,
3]. For stage IB patients, different guidelines give discordant recommendations [
2,
3]. Therefore, it is important to find a clinically actionable biomarker for the prediction of prognosis and adjuvant chemotherapy benefits in patients with resected stage I NSCLC.
Computed tomography (CT) is routinely used in lung cancer diagnosis and could provide the possibility of calculating prognostic and predictive biomarkers for patients’ management. CT-based radiomics has become an attractive method for predicting gene mutations, treatment sensitivity, and prognosis in NSCLC [
4‐
6]. Recently, Xie et al. indicated that the radiomics nomogram could be used for prognostic prediction (hazard ratio [HR] = 7.794; 95% confidence interval [CI]: 3.185–19.078;
P < 0.001) and adjuvant chemotherapy benefits identification for patient with resected stage I lung adenocarcinoma (
P = 0.040) [
7]. However, only disease-free survival (DFS) was assessed in that study.
The goal of cancer treatment is to improve patient survival. Overall survival (OS) is an optimal end point which indicates patient benefit. In the present study, we developed and validated a radiomic signature (RS) to predict OS in lobectomized stage I lung adenocarcinoma, and then set out to explore if the RS could be a biomarker for identification patients who can benefit from adjuvant chemotherapy.
Methods
Study design and patient population
This was a retrospective multicenter study including three independent cohorts (n = 474, Additional file
1: Figure S1). The training cohort and internal validation cohort included 287 and 122 patients who were treated in Shanghai Chest Hospital from March 2009 to December 2013 and January 2014 to May 2016, respectively. The external validation cohort included 65 patients who were treated in Xinhua Hospital from February 2015 to March 2021.
All patients diagnosed with stage I lung adenocarcinoma according to the 8th TNM edition of the American Joint Committee on Cancer (AJCC) cancer staging manual were identified [
8]. Patients were excluded based on the following criteria: (1) multiple primary lung adenocarcinoma; (2) adenocarcinoma in situ and microinvasive adenocarcinoma (MIA); (3) without high resolution CT images carried out less than 2 weeks before surgery; (4) died within 1 month after surgery; (5) lost to follow-up.
The demographic and clinicopathological characteristics were collected from medical records databases in two hospitals. Platinum-based doublet chemotherapy was used as the basic regimen. The platinum drugs included cisplatin and carboplatin, and other drugs included vinorelbine, paclitaxel, gemcitabine or pemetrexed.
Ethics committee approval was obtained from the institutional review boards at Shanghai Chest Hospital (KS1596, 08/28/2019) and Xinhua Hospital (XHEC-D-2021–137, 10/18/2021). Informed consent was waived because data were deidentified.
Outcome and follow-up
The primary endpoint was OS, which was defined as the time from surgery to death from any cause. All patients were postoperatively followed every 3 months during the first 2 years and then every 6 months annually thereafter. The clinical evaluations included physical examination, blood tests, chest CT, abdominal CT, or ultrasound. Whole-body bone scans and a cranial CT or magnetic resonance imaging were performed annually.
CT data acquisition and imaging feature detection
Chest CT scans were performed using the following 4 scanners: Discovery CT750HD CT scanner (GE, Waukesha, WI, USA), 256-detector row scanner (Revolution CT, GE, Waukesha, WI, USA), 64-detector row scanner (Brilliance, Philips, Cleveland, OH, USA), and a 16-detector row scanner (uCT S160, United Imaging, Shanghai, China). Patients were scanned at the end of inspiration during a single breath hold in the supine position. The HRCTs were performed with collimation of 0.625–1.25 mm, pitch of 0.64, section thickness of 0.625–1.25 mm without overlap, matrix of 512 × 512 or 1,024 × 1,024, field of view (FOV) of 350–400 mm, 120 kVp, and 220–300 mA. All imaging data were reconstructed using the standard algorithm. One radiologist (G.T., with 10 years of experience in chest CT interpretation) identified manually at the voxel level the areas of interest for the included nodules based on CT scans using 3D Slicer (version 4.8.0, Brigham and Women’s Hospital, Boston, MA, USA). Then, the VOI was confirmed by another radiologist (H.Y., with 30 years of experience in chest CT interpretation).
Radiomics features were extracted by Image Biomarker Standardization Initiative (IBSI) compliant AK software (Analysis Kit Software, Version3.3.0, GE Healthcare). Totally, 1218 radiomic features were extracted from CT images, including first order statistical features, morphological features, gray-level co-occurrence features matrix-based features, gray-level run length matrix-based features, gray-level size zone matrix-based features, gray-level dependence matrix-based features, and the transform features of wavelet and Laplace changes.
Radiomic signature construction and validation
The RS was calculated with chest CT based on the training cohort. Univariate Cox analysis was firstly used to detect the associations between each feature and the patients’ OS. The features with
P < 0.05 were used for further analysis. The Spearman correlation was applied to eliminate the redundancy of the feature set (coefficient of chosen here | r |> 0.8). Finally, the least absolute shrinkage and selection operator (LASSO) method to select the most valuable prognostic features from the training cohort. The optimal cutoff value for RS was determined using X-tile software version 3.6.1 (Yale University School of Medicine, New Haven, CT, USA) in the training cohort [
9]. The same cutoff value was applied to all the validation cohorts. The patients were divided into high and low risk groups in each cohort. The time dependent receiver operating characteristic (ROC) curve was created to assess the prognostic accuracy of the RS in the training and two validation cohorts. To adjust for selection bias, propensity score matching (PSM) was did, due to the imbalanced data between low and high risk groups. The propensity score was assessed for training and internal validation sets [
10]. Nearest neighbor matching was selected and matched controls were not replaced during matching. Using the propensity scores, high risk patients were randomly matched to low risk patients with 1:1 matching method.
Statistical analysis
Mann–Whitney U test was used to examine the difference between the two groups. Categorical data were compared using the χ2-test or Fisher’s exact test, as appropriate. OS was calculated using the Kaplan–Meier method and log-rank test. The univariate and multivariate Cox proportional hazards model was utilized to estimate the HR and 95% CI for the outcome. Interaction between the RS and adjuvant chemotherapy was assessed by means of the Cox model.
We established a clinicopathologic model and a radiomic nomogram to determine whether the RS added incremental value for predicting OS. Model performance was assessed by Harrell’s concordance index (C-index), calibration curves and decision curve analysis.
For all analyses, P < 0.05 was considered statistically significant in all 2-tailed tests. The statistical analyses were performed using R version 3.6.1 (R Project for Statistical Computing) and SPSS version 23.0 (IBM, Armonk, NY).
Discussion
In the present study, we developed the RS by preoperative CT images and validated its ability to predict OS in three cohorts from two centers. More importantly, the RS might predict response to adjuvant chemotherapy in lobectomized stage I lung adenocarcinoma, especially in stage IB.
In a previous study, Xie and colleagues found that RS could predict RFS and identify the patients benefit from adjuvant chemotherapy in stage I adenocarcinoma patients (
P = 0.040) [
7]. Most studies about postoperative adjuvant chemotherapy used OS as the primary endpoint [
11‐
13]. In an individual participant data meta-analysis, Burdett et al. determined a benefit of adding chemotherapy after surgery (HR = 0.86; 95% CI: 0.81–0.92;
P < 0.001), with an absolute increase in OS of 4% at five years [
12]. In addition, after analyzing 577 NSCLC patients from two data sets, Le et al. indicated that the risk score generated using CT-based radiomics signatures could predict overall survival in NSCLC patients [
14]. Thus, we focused on OS in this study and also revealed RS's prognostic and predictive potential. We found that high RS was associated with solid/micropapillary subgroup and tumor size, which was consistent with the previous report [
7]. When PSM was used to minimize potential selection bias and confounding effects, RS could also predict OS. Therefore, pathological subtypes and tumor size may not be the explanations for the relationship between RS and OS. Perez-Johnston et al. suggested that other clinicopathological and genomic features, such as tumor spread through air spaces, phosphoinositide 3-kinase pathway or
STK11 alterations, were enriched in certain CT-based radiomic clusters [
15]. Therefore, it was possible that these clinicopathological and genomic features might influence the association between RS and OS.
In patients with stage IA, high RS was associated with worse OS. However, adjuvant chemotherapy could not improve survival in these high risk patients. Therefore, clinicians should formulate a detailed follow-up plan in order to detect local or metastatic relapse. Molecular residual disease (MRD) detection could precisely predict the recurrence in patients with NSCLC after definitive surgery [
16]. It will be interesting to use MRD detection to predict survival in high risk patients with stage IA.
Poorly differentiated, lymphovascular invasion, visceral pleural invasion, incomplete lymph node sampling, or wedge resection were defined as high risk factors in stage IB by the current NCCN guideline [
3]. Several studies reported stage IB patients with these high risk factors derived survival benefit from adjuvant chemotherapy [
17,
18]. However, it is still controversial to identify patients with early-stage NSCLC who may benefit from adjuvant chemotherapy after surgery. In our study, improved OS was observed in adjuvant chemotherapy receivers with a high RS, suggesting its ability to predict survival and response to chemotherapy. However, the sample size of patients receiving adjuvant chemotherapy was moderate, which might limit the statistical power of conclusions. Thus, the radiomic signature as a noninvasive method should be assessed in the large-scale prospective studies.
Osimertinib has been recommended to use in patients with stage IB
EGFR mutation-positive NSCLC [
19]. In addition, radiomic signature may predict the prognosis of metastatic NSCLC patients with receiving osimertinib therapy [
20]. Therefore, the association between the radiomic signature and efficacy of osimertinib as adjuvant therapy is needed to be investigated.
There were some limitations in our study. First, as a retrospective study, potential selection bias may hamper the reproducibility and comparability of the results. Thus, we included three independent cohorts from two medical centers to validate our findings. Second, genetic data was not included because gene detection was not a routine practice. Third, the role of radiomic signature was only assessed in Chinese patients. The performance of RS in other ethnic patients was still unknown. Finally, it would be interesting to perform experiments with cell lines and animal models to reveal the underlying mechanism of radiomic signature.
Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit
http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (
http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.