Abstract
Background
Wide use of low-dose chest CT detects more early NSCLC, though recurrence (mostly extracranial distant metastasis) still occurs. This study aimed to establish and validate computed tomography (CT) radiomics models that can be used for predicting the risk of extracranial distant metastasis (EDM) after radical resection in early-stage non-small cell lung cancer (e-NSCLC).
Methods
This retrospective study included 312 patients with e-NSCLC who underwent radical surgery at Shandong Cancer Hospital from 2014 to 2019. The cases were randomly divided into a training set (n=200) and a validation set (n=112). The study excluded patients who received neoadjuvant therapy, had interrupted follow-up, or died of non-tumor-related causes. The patients were followed up for an average of 30.6 months after the first postoperative CT examination. This study defined EDM as distant metastasis in other organs other than brain metastasis after surgery, and accounted for 78.6% of all metastatic cases. The region of interest (ROI) was delineated, and the radiomics features were selected by using least absolute shrinkage and selection operator (LASSO) logistic regression models to construct radiomics signatures. The association between the radiomics signature and the occurrence of EDM was also evaluated. A combined nomogram incorporating the selected radiomics and clinical-pathological predictors was constructed using multivariable regression. The concordance index (C-index) was used to evaluate the performance of the model. To determine the clinical effectiveness of nomogram in radiomics, the decision curve analysis (DCA) was performed.
Results
A total of 75 patients (24.0%) experienced metastasis, with EDM accounting for 78.6%. In the overall cohort, radiomics signature were significantly associated with the occurrence of the EDM. The discrimination performance of the radiomics nomogram is significantly better than the clinical-pathologic nomogram (P<0.001), with the C-indexes 0.818 [95% confidence interval (CI): 0.780–0.856] and 0.823 (95% CI: 0.769–0.877) in the training and validation cohorts, respectively. Analysis of the calibration and decision curves indicated that the comprehensive nomogram was superior to the clinical-pathological and radiomics nomogram in predicting EDM survival.
Conclusions
CT radiomics features have the potential to predict postoperative EDM and may assist in preliminarily identifying patients at high metastatic risk, which could provide a reference for individualized postoperative adjuvant decision-making in e-NSCLC.
Keywords: Early-stage non-small cell lung cancer (e-NSCLC), distant extracranial metastases, radiomics, nomogram
Highlight box.
Key findings
• Extracranial distant metastasis (EDM) is an important recurrence pattern after radical resection in early-stage non-small cell lung cancer.
What is known and what is new?
• Computed tomography-based comprehensive radiomics nomogram was constructed to evaluate the risk of EDM.
• The efficacy of the integrated nomogram was superior to that of the clinical-pathological and radiomics signature in predicting EDM survival.
What is the implication, and what should change now?
• CT imaging features could be used to predict the occurrence of EDM in early-stage NSCLC patients after radical resection and may be helpful for identifying high-risk groups for postoperative adjuvant treatments.
• Analysis of patients in the high-risk group assessed by the combined nomogram showed that patients who had previously received postoperative adjuvant chemotherapy had a survival benefit compared with patients who had not.
Introduction
With the widespread implementation of low-dosage chest computed tomography (CT) screening for lung cancer, early-stage non-small cell lung cancer (e-NSCLC) has been increasingly detected worldwide (1,2). Surgical procedures have been the first-choice, standard care for e-NSCLC. However, an inherent risk of recurrence and metastasis persists, even with a 5-year survival rate ranging from 65% to 90% for patients with stage IA-IIA disease (3,4). Distant metastasis is the main model of recurrence after radical resection in e-NSCLC, with extracranial distant metastasis (EDM) accounting for the greatest proportion (74–89%) (5,6).
Currently, the tumor, node, metastasis (TNM) staging system is the most widely accepted method for predicting the postoperative outcomes in patients with e-NSCLC (3). Due to tumor heterogeneity, even lung cancer patients who receive similar treatment and have the same TNM stage may have different survival outcomes. The clinical predictive indicators for distant metastasis after surgery for e-NSCLC have certain limitations, and there are no relevant factors for predicting EDM (7,8). This model is designed to assist two types of clinical decisions: postoperative follow-up planning and joint decision making for adjuvant therapy (9,10). After the patient undergoes radical surgical resection, the preoperative CT radiomics features and postoperative clinical pathological variables are integrated to stratify the patients (11). For those patients determined to have a high risk of recurrence, enhanced follow-up can be considered, including shortening the interval between CT examinations (5) and weighing the potential benefits of adjuvant therapy (12,13). In contrast, low-risk patients can adopt a more lenient follow-up plan to avoid unnecessary medical burden (9). It is expected to become a practical tool for individualized postoperative management of e-NSCLC (9).
Several studies suggest that radiomics analysis could enhance the efficiency of recurrence prediction in patients with e-NSCLC (14,15). At present, radiomics has a high application value in predicting distant metastasis in early-stage lung cancer (16). Most existing models have a wide range of enrolled populations, typically including all postoperative cases from stages I to III. Some patients also received neoadjuvant therapy, and the homogeneity of the population needs to be improved (12,13). In addition, although large sample models constructed based on databases have a large sample size, they lack consideration of real-world imaging scanning conditions and pathological details. Due to the significant differences in biological behavior and clinical decision-making between intracranial metastasis and extracranial systemic metastasis, setting extracranial metastasis as the research endpoint can ensure the homogeneity of the research endpoint and its clinical relevance (17). However, to our knowledge, the use of CT-based radiomics to predict early distant metastasis (EDM) in patients undergoing surgery for e-NSCLC has not been previously documented.
In this study, we aimed to examine the possibility of using radiological and clinicopathological features to predict postoperative EDM of e-NSCLC. In addition, we established a comprehensive radiological and clinicopathological nomogram. This study makes it possible to further stratify the risk of such patients. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1263/rc).
Methods
Patients
Patients with early-stage (IA–IIA) NSCLC from January 1, 2014, to December 30, 2019, were retrospectively identified from the medical records database of Shandong Cancer Hospital. This study inclusion criteria included patients who were admitted and identified as patients with e-NSCLC who underwent radical resection, had complete follow-up information and underwent preoperative enhanced CT. Patients with at least one of the following were excluded from the study: (I) patients who received other treatments before the surgery and (II) patients who died of other causes or lacked follow-up information and baseline data. Details of patient recruitment pathways, as well as inclusion and exclusion criteria, can be found in Figure S1. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective study was exempted from ethical review, and informed consent was waived owing to the use of de-identified data. Tumor staging was performed for enrolled patients in accordance with the TNM Joint Committee’s eighth edition of the Staging Manual (3). All patients were assigned in a 2:1 ratio to the training group and the validation group using computer-generated random numbers.
Follow-up
In this study, EDM was defined as distant metastasis excluding brain metastasis after surgery, meaning metastasis to other organs, such as the liver, bone, peritoneum, or adrenal gland, during the follow-up period. EDM was confirmed by imaging examinations [e.g., CT, positron emission tomography-computed tomography (PET-CT)] or pathological biopsy when feasible, with brain metastasis explicitly excluded. EDM-DFS was defined as the time interval from the start date of treatment to the date of the first appearance of EDM. According to our hospital’s follow-up plan, follow-up with chest enhanced CT was performed every 3 months, 6 months, and annually after resection. Baseline clinicopathological data were acquired from the medical record system, including age, sex, smoking status, ECOG (Eastern Cooperative Oncology Group) score standard, clinical stage, tumor differentiation degree, tumor histology type, tumor pathological diameter, visceral pleural invasion status, tumor location (peripheral or central), adjuvant chemotherapy (ACT) and baseline CT imaging data. The median postoperative follow-up time was 30.6 months (16.6–42.5 months) after the first CT examination. The follow-up deadline was June 1, 2020. The criteria for determining the end point of the follow-up as censored were: (I) no distant metastasis outside the brain occurred by the end of the follow-up; (II) the patient was lost to follow-up during the follow-up process; (III) non-tumor-related death occurred. For cases where death occurred during the follow-up period but there was no imaging evidence of distant metastasis outside the brain, they will be censored at the time of death.
Image capture and radiomics features extraction
All participants underwent preoperative chest CT scans with contrast enhancement, from which the CT images were acquired. Using a 64-detector scanner (SOMATOM Definition AS+, Siemens). Scanning parameters were set as follows: tube voltage of 120 kVp, tube current of 160 mA, and reconstructed slice thickness of 5 mm. Images were viewed with lung window (window width 1,600 HU, window level −600 HU) and mediastinal window (window width 400 HU, window level 40 HU). Venous-phase CT images were exported in DICOM format for radiomic feature extraction. The areas of interest were manually delineated by two blinded, experienced radiologists (with extensive clinical experience, especially in CT image analysis and interpretation). A total of 850 CT image characteristics were extracted, including original image shape, wavelet, gray-level co-occurrence matrix (GLCM), first-order, gray run length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighbourhood gray difference matrix (NGRLM) (18). The intensity characteristics were calculated with the first-order statistics of the tumor voxel intensity distributions. The surface area, compactness, and tumor volume were extracted to reflect the three-dimensional geometric characteristics of the tumor. Texture matrices such as the grey-level co-occurrence matrix (GLCM) and grey-level size zone matrix (GLSZM) were used to quantify internal tumor heterogeneity, and log filters were used to depict tumor coarseness with different sigma values to describe texture features. Wavelet decomposition of the original image was used to extract intensity and texture features in the frequency domain (19).
Statistical analysis
Descriptive statistics were used to describe the characteristics of each cohort participant, the mean and standard deviation were used for continuous variables, and frequencies and percentages were used for classification variables. To evaluate the training and validation cohorts, we observed the difference in clinicopathological baseline features between the two groups by using χ2 test, an independent samples t-test, Mann-Whitney U test. We calculated the interclass correlation coefficients (ICCs) to evaluate the reproducibility of observer measurements for key features. Furthermore, an ICC value above 0.8 was considered indicative of good consistency.
The least absolute shrinkage and selection operator (LASSO) method was used to pick out the most helpful predictive signatures from the original data set. The radiomic score for each patient was weighted by its respective coefficients using a linear combination of the selected signatures. Kaplan-Meier survival analysis was applied to assess the correlation among the radiomics features and EDM. The rad-scores were used to divide objects into high-risk and low-risk sets, and the threshold was determined by applying X-tile. By considering radiomics markers to be independent biomarkers, clinicopathological hazards were integrated into multivariate Cox proportional risk factor models. According to the results of Cox multivariate analysis, the radiomics signature model and clinicopathological model were constructed respectively, as well as the combined model including clinical independent risk factors and radiomics features, namely the radiomics model. The concordance index (C-index) was used to evaluate the efficiency of the model (20).
To assess the incremental value of the radiomics signature, a radiomics nomogram and a clinicopathological nomogram of the validation group were constructed. A calibration curve was developed to contrast the conformance of the observations with the relevance between the patterns of extracranial metastasis failure in the radiomics and clinicopathological nomograms (21). Bootstrap validation (1,000 resamples) was performed to compute the bias corrected C-index within the validation cohort. Cox regression was employed and the proportional hazards assumption was tested. To determine the clinical effectiveness of the radiomics nomogram, decision curve analysis (DCA) was performed to quantify the net benefit at diverse threshold probabilities (22).
To further verify whether the selected high-risk population with EDM could benefit from postoperative ACT, the total point of the nomogram for each patient was calculated, and the risk grade was determined. In addition, the benefit of ACT in high-risk patients screened by comprehensive prediction model was retrospectively analysed. Kaplan-Meier survival curves were constructed for both groups to evaluate treatment effectiveness.
Statistical analyses of this study were performed with IBM SPSS (version 24.0, IBM SPSS), X-tile software, version 3.6.1 (Yale University School of Medicine, New Haven, Conn), R software, version 3.6.1 (http://www.R-project.org) and MedCalc (version 15.6.1, MedCalc) software. The significant level for the statistics was set at 0.05.
Results
Clinicopathologic features of the patients
In this study, 312 patients were enrolled (188 men and 124 women; mean age, 63.23±8.234 years; range, 32–84 years). Among the 312 N0-NSCLC patients, the distribution of TNM stages was as follows: IA, 221 patients (70.8%); IB, 63 patients (20.2%); and IIA, 28 patients (9.0%)—details are presented in Table 1.There was no difference between the two groups in either clinical features or follow-up information (P=0.059–0.966; Table 1). In total, 75 (24.0%) patients experienced disease recurrence after curative surgical treatment, and the failure pattern is shown in Table 2. EDM accounted for a higher proportion of patients with recurrence. In our study, the incidence of EDM was as high as 78.6%. Lung cancer caused the death of 12 patients (4.5%). The median extracranial disease-free survival (DFS) was 30.0 months, and the 25th and 75th percentiles were 14.5 and 41.9 months, respectively.
Table 1. Clinical parameters of the training cohort and validation cohort.
| Clinical parameter | All patients (n=312) | Training cohort (n=200) | Validation cohort (n=112) | P value |
|---|---|---|---|---|
| Age (years) | 62.2±8.2 | 62.3±8.3 | 62.2±8.2 | 0.85 |
| Gender | 0.19 | |||
| Male | 188 (60.3) | 126 (63.0) | 62 (55.4) | |
| Female | 124 (39.7) | 74 (37.0) | 50 (44.6) | |
| Smoking status | 0.34 | |||
| Never | 156 (50.0) | 114 (57.0) | 42 (37.5) | |
| Frequently | 156 (50.0) | 86 (43.0) | 70 (62.5) | |
| ECOG | 0.26 | |||
| 0 | 229 (73.4) | 151 (75.5) | 78 (69.6) | |
| ≥1 | 83 (26.6) | 49 (24.5) | 34 (30.4) | |
| Tumor diameter (pathologic) (cm) | 0.71 | |||
| Mean (range) | 2.6 (0.6–5.0) | 2.5 (0.6–5.0) | 2.6 (0.7–5.0) | |
| ≤3 cm | 221 (70.8) | 145 (72.5) | 76 (67.9) | |
| >3 cm | 91 (29.2) | 55 (27.5) | 36 (32.1) | |
| Pleural invasion | 0.97 | |||
| Yes | 137 (43.9) | 88 (44.0) | 49 (43.8) | |
| No | 175 (56.1) | 112 (56.0) | 63 (56.2) | |
| Clinical stage (AJCC v.8) | 0.77 | |||
| IA | 221 (70.8) | 142 (71.0) | 79 (70.5) | |
| IB | 63 (20.2) | 43 (21.5) | 20 (17.9) | |
| IIA | 28 (9.0) | 15 (7.5) | 13 (11.6) | |
| Location | 0.63 | |||
| Peripheral | 206 (66.0) | 134 (67.0) | 72 (64.3) | |
| Central | 106 (34.0) | 66 (33.0) | 40 (35.7) | |
| Tumor histology | 0.29 | |||
| SCC | 68 (21.8) | 40 (20.0) | 28 (25.0) | |
| AC | 239 (76.6) | 156 (78.0) | 83 (74.1) | |
| Others | 5 (1.6) | 4 (2.0) | 1 (0.8) | |
| Histological grade | 0.77 | |||
| Well differentiated | 14 (4.5) | 9 (4.5) | 5 (4.5) | |
| Moderately differentiated | 242 (77.6) | 154 (77.0) | 88 (78.5) | |
| Poorly differentiated | 56 (18.0) | 37 (18.5) | 19 (17.0) | |
| Adjuvant chemotherapy | 0.61 | |||
| Yes | 150 (48.1) | 94 (47.0) | 56 (50.0) | |
| No | 162 (51.9) | 106 (53.0) | 56 (50.0) | |
| Follow-up time (months) | 0.059 | |||
| Median(IQR) | 30.6 (16.6–42.5) | 29.0 (15.0–42.1) | 32.5 (19.4–44.1) | |
| Maximum | 72.2 | 72.2 | 71.4 | |
| Relapse state (EDM) | 0.72 | |||
| 0 | 253 (81.1) | 161 (80.5) | 92 (82.1) | |
| 1 | 59 (18.9) | 39 (19.5) | 20 (17.9) |
Data are presented as n (%) or mean ± standard deviation unless otherwise specified. There was no difference between the training cohort and the validation cohort in either the clinical-pathological characteristics or the follow-up time (P=0.06–0.97). Clinical-pathological variables of training cohort and validation cohort: the data are the number of patients and the data in parentheses are percentages unless otherwise specified. AC, adenocarcinoma; AJCC, American Joint Committee on Cancer; ECOG, Eastern Cooperative Oncology Group score standard; EDM, extracranial distant metastasis; IQR, interquartile range; SCC, squamous cell carcinoma.
Table 2. Failure patterns for all patients.
| Failure patterns | Training data [n=49 (65.3%)] | Validation data [n=26 (34.7%)] | Total (n=75) |
|---|---|---|---|
| Single | 36 | 19 | 55 |
| Multiple | 13 | 7 | 20 |
| Local + regional lymph node recurrence | 12 | 10 | 22 |
| Distant recurrence | 44 | 23 | 67 |
| Extracerebral distant metastasis | 39 | 20 | 59 |
| Intrathoracic distant metastasis | 18 | 7 | 25 |
| Bone metastasis | 12 | 5 | 17 |
| Hepatic metastasis | 5 | 2 | 7 |
| Adrenal gland metastasis | 5 | 2 | 7 |
| Kidney metastasis | 1 | 0 | 1 |
| Colon metastasis | 1 | 0 | 1 |
| Spleen metastasis | 1 | 0 | 1 |
| Brain metastasis | 5 | 3 | 8 |
In all recurrence cases, 59 cases (18.9%) were extracerebral distant metastases, including 25 intrathoracic (distant) metastases, 17 bone metastases, 7 hepatic metastases, 7 adrenal gland metastases, 1 kidney metastasis, 1 colon metastasis, 1 spleen metastasis.
Establishment and verification of the radiomics score
Within the training data, enhanced CT images were obtained, and 850 radiomics signatures were extracted; they could be divided into 7 categories: original and wave processed shape, GLDM, GLCM, FirstOrder, GLRLM, GLSZM, and NGTDM. Then, a top signature subset building of 6 radiomics features that were the most strongly related to DFS was identified by LASSO logistic regression analysis (Figure 1). The AUC value of the training group was 0.784 and that of the verification group was 0.762. CT imaging features were predictive of EDM occurrence with certain accuracy in both the training and validation datasets. Subsequently, the radiomics score (rad-score) was established based on the formula generated with the LASSO model. The rad-score is calculated by multiplying the total of the values of each factor by its modulus:
Figure 1.
The ROC curve for the radiomics features. The AUC value of the training group was 0.784 and that of the validation group was 0.762. A LASSO binary logistic regression model was used to select radiomics features. The AUC curve was drawn versus log(λ). A λ-value of 0.138997 with log (λ)=−1.973303 was selected according to 10-fold cross-validation. A vertical line is drawn after 10-fold cross-validation of the selected values, resulting in six nonzero coefficients. AUC, area under the curve; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic.
| [1] |
The optimal cut-off number produced by the X-tile plot was 1.2 (Figure S2). Accordingly, all of the patients were divided into two groups: the high-risk cohort (rad-score ≥1.2) and the low-risk cohort (rad-score <1.2). The Kaplan-Meier curves for the training and verification groups showed that the survival rate of the low-risk data set was above that of the high-risk data set, and the variation was statistically significant (P<0.001, Figure 2A,2B). Objects with higher rad-scores usually had shorter EDM-DFS. Extracranial failure was identified in 43 patients (72.9%) in the higher-risk cohort (mean, 26.3 months) but only 16 patients (27.1%) in the lower-risk cohort (mean, 33.5 months).
Figure 2.
Kaplan-Meier survival analyses on the basis of the radiomics features for patients in the training group (A) and those in the validation group (B).
Construction and validation of nomograms
Cox univariate regression analysis revealed that 5 predictors, including 4 clinicopathological predictors (tumor pathological diameter, ECOG, smoking history, tumor differentiation degree) and the rad-score, were correlated with EDM. Multivariate analysis showed that tumor pathological diameter, smoking history, ECOG and rad-score showed relatively better prognostic ability for EDM, with HRs of 1.348 [95% confidence interval (CI): 1.006 to 1.806; P=0.045], 1.953 (95% CI: 1.122 to 3.400; P=0.02), 0.471 (95% CI: 0.271 to 0.819; P=0.008), and 3.338 (95% CI: 1.703 to 6.541; P<0.001, Table 3), respectively.
Table 3. Clinical-pathological univariate and multivariate Cox proportional hazards regression modeling for extracerebral distant failure.
| Parameter | Univariate analysis | Multivariate analysis | |||
|---|---|---|---|---|---|
| Hazard ratio (95% CI) | P value | Hazard ratio (95% CI) | P value | ||
| Age | 0.996 (0.966–1.027) | 0.80 | |||
| Sex (M vs. F) | 1.141 (0.673–1.935) | 0.62 | |||
| Smoking history (yes vs. no) | 2.772 (1.635–4.702) | <0.001* | 1.953 (1.122–3.400) | 0.02* | |
| ECOG (0 vs. ≥1) | 0.331 (0.199–0.552) | <0.001* | 0.471 (0.271–0.819) | 0.008* | |
| Tumor histology (AC vs. SCC) | 0.946 (0.518–1.730) | 0.86 | |||
| Visceral pleural invasion (yes vs. no) | 1.530 (0.910–2.572) | 0.11 | |||
| Tumor differentiation degree | 0.002* | 0.28 | |||
| High vs. low | 0.385 (0.226–0.656) | <0.001* | 0.633 (0.351–1.142) | 0.13 | |
| Moderate vs. low | 0.324 (0.076–1.378) | 0.13 | 0.507 (0.116–2.218) | 0.37 | |
| Tumor diameter (pathologic) | 1.802 (1.393–2.332) | <0.001* | 1.348 (1.006–1.806) | 0.045* | |
| Clinical stage (AJCC v.8) | 0.16 | ||||
| IB vs. IA | 1.660 (0.930–2.965) | 0.09 | |||
| IIA vs. IA | 1.643 (0.729–3.702) | 0.23 | |||
| Tumor location (central vs. peripheral) | 1.084 (0.636–1.847) | 0.77 | |||
| Radiomics score | 7.768 (4.226–14.279) | <0.001* | 3.338 (1.703–6.541) | <0.001* | |
The table presented clinical-pathological features and radiomics score (rad-score) into univariate survival analysis, which included age, sex, smoking history, ECOG, tumor histology, visceral pleural invasion status, tumor differentiation degree, tumor diameter (pathological), clinical stage, tumor location (peripheral or central). Subsequently, the univariate survival analysis revealed 5 predictors, including 4 clinical-pathological (tumor pathological diameter, ECOG, smoking history, tumor differentiation degree) and rad-score, were correlated to the EDM. Multivariate survival analysis showed that tumor diameter (pathological), ECOG, smoking history and rad-score were associated with EDM. *, P<0.05. AC, adenocarcinoma; AJCC, American Joint Committee on Cancer; CI, confidence interval; ECOG, Eastern Cooperative Oncology Group; EDM, extracranial distant metastasis; F, female; M, male; Rad-score, radiomics score; SCC, squamous cell carcinoma.
Evaluating additional value in individualized prediction of radiomics features
The radiomics nomogram, clinicopathological nomogram and rad-score nomogram are shown in Figure 3A. The C-index of the clinicopathological nomogram was 0.778 (95% CI: 0.735–0.822) in the training group. The C-index of the rad-score nomogram was 0.781 (95% CI: 0.734–0.829) in the training group. The C-index of the radiomics nomogram was 0.818 (95% CI: 0.780–0.856) in the training group. The C-index of the clinicopathological nomogram was 0.746 (95% CI: 0.682–0.810) in the validation data set. The C-index of the rad-score nomogram was 0.754 (95% CI: 0.684–0.825) in the validation data set. The C-index of the radiomics nomogram was 0.823 (95% CI: 0.769–0.877) in the validation cohort (Table 4, Figure S3). To evaluate the efficacy of predicting the incidence of EDM at 12, 24, and 36 months after resection, calibration curves for the nomograms were plotted (Figure 3B-3D) and indicated that the assessment of the radiomics nomogram and realistic observation had nice consistency. DCA was applied to identify the net clinical benefit, and it was evident that the radiomics nomograms were better than the clinicopathologic nomograms (Figure 3E).
Figure 3.
The comprehensive nomogram. (A) The nomogram was obtained by quantifying the risk factors identified by Cox regression analysis in the training group. The corresponding values of risk factors for patients are added to predict the probability of extracranial distant metastasis at 1, 2 and 3 years. (B-D) The calibration curves of the 1-, 2- and 3-year radiomics nomograms are shown in the figure, demonstrating the consistency between the estimated results and the observed results. (E) The figure shows the decision curve analysis for each model. It can be seen from the figure that the combined radiomics nomogram is significantly better than the other prediction models. ECOG, Eastern Cooperative Oncology Group; EDM, extracranial distant metastasis.
Table 4. Comparison of C-index between different nomograms of training group and validation group.
| Model type | Training group | P value | Validation group | P value |
|---|---|---|---|---|
| 1 | 0.818 (0.780–0.856) | <0.001 (1 vs. 2) | 0.823 (0.769–0.877) | <0.001 (1 vs. 2) |
| 2 | 0.778 (0.735–0.822) | <0.001(1 vs. 3) | 0.746 (0.682–0.810) | <0.001(1 vs. 3) |
| 3 | 0.781 (0.734–0.829) | 0.74 (2 vs. 3) | 0.754 (0.684–0.825) | 0.66 (2 vs. 3) |
Data are presented as C-index (95% CI). Model 1: comprehensive radiomics nomogram; Model 2: clinical-pathologic nomogram; Model 3: rad-score nomogram. CI, confidence interval.
To verify the benefit of postoperative ACT in patients with high scores according to the radiomics nomogram, previously treated patients were analysed, and the results revealed that ACT was statistically significant in the high-score group (total points ≥4.1, P=0.003) but not in the low-score group (total points <4.1, P=0.07, Figure S4).
Discussion
This study found a high incidence of EDM after surgery for e-NSCLC, which may affect long-term survival postoperatively. Therefore, a comprehensive nomogram based on a radiomics feature for individualized, preoperative prediction of EDM in patients with e-NSCLC was developed and validated. We found that the predictive value of radiological features combined with the clinicopathologic risk nomogram was superior to that of a single factor.
Radiomics involves analysing a large amount of quantitative data to characterize an algorithm and comprehensively quantify the phenotype of a tumor (23,24). Currently, radiomics is advancing swiftly within oncology, with a growing number of predictive models being validated (25). It can be seen from the characteristics screened in this study, the intensity characteristics were calculated with the first-order statistics of the tumor voxel intensity distributions and wavelet was used to extract intensity and texture features in the frequency domain. In some recent studies, it has been demonstrated that radiomics signatures can be used to stratify patients according to different levels of risk, better predict patient outcomes and enable treatment optimization for patients (26,27). Therefore, we validated the radiomics signature as an independent prognostic factor and successfully classified objects into high-risk and low-risk data sets. Risk stratification aids clinicians in early adjustment of treatment plans, thereby enhancing patient prognoses. At present, radiomics has good predictive value in predicting recurrence and metastasis in e-NSCLC and the occurrence of brain metastasis (15,16,28). Radiomics also provides new insights for predicting distant metastasis in stage I NSCLC (16). To date, there has been no report on the use of CT radiomics to predict the risk of EDM in e-NSCLC.
Many studies have combined radiomics features and clinicopathological factors associated with survival and prognosis in nomograms to better visualize predicted therapeutic effects. A comprehensive radiomics nomogram had a better ability to predict the recurrence of lung cancer than single radiomics feature models or clinical models, which has been confirmed in previous studies (29). Huang et al. investigated whether combining the radiomics features into a radiomics-based nomogram contributed to a higher assessment of DFS than that achieved with a clinicopathological nomogram and other aspects, such as nicer calibration and improved precision in classifying survival outcomes (15). Some research data have indicated that clinical factors alone are quite impartial in their ability to predict metastasis. This may be related to the specific treatment plan, the patient’s physical condition and genotype, and other social factors, such as family, economic and psychological status. The development of a stable and effective prediction model is essential. To date, few studies have reported on the prediction of EDM in e-NSCLC, especially models that include all predictive indicators.
To improve the postoperative survival rate in lung cancer patients, adjuvant therapy is generally performed postoperatively to reduce recurrence and metastasis. There is no evidence that postoperative radiotherapy improves survival in patients with lung cancer (stage I–II), and the side effects of radiotherapy may increase the risk for patients (30). Molecularly targeted therapy for lung cancer is commonly used in advanced NSCLC, but how targeted therapy can be combined with surgery to improve survival in T1a-2aN0M0 NSCLC requires additional high-quality research (31,32). ACT is currently the most widely used adjuvant treatment, and the benefits and risks of ACT need to be comprehensively evaluated before applying ACT after complete tumor resection in NSCLC patients due to the substantial side effects of chemotherapy drugs. Patients with stage IA NSCLC after complete resection should be followed up regularly, and ACT is not recommended (33). The latest NCCN guidelines recommend ACT for patients with stage IB-IIA NSCLC after surgery in the manifestation of high-risk factors such as vascular invasion, wedge resection, tumors >4 cm, poorly differentiated tumors, visceral pleural involvement and unknown lymph node status (NX) (34). Studies have shown benefits from ACT for high-risk patients with EDM after surgery for e-NSCLC, which was also the major pattern of recurrence (5,6). Based on these studies, we combined a radiomics signature with clinicopathological factors to create a simple and intuitive nomogram. This preprocessing CT-based radiomics model may be capable of predicting EDM of e-NSCLC and makes it possible to identify high-risk patients from the NSCLC population who would benefit from postoperative adjuvant therapy in advance.
There were some limitations in this study. First, this study was a retrospective study with data from a single study centre, so patients’ management pathways were roughly similar. Second, the radiomics feature extraction was based on the delineated target area. The experience of the radiologist in actual clinical application has a certain influence on the results of the experiment. Previous studies have found that 3D slicer semiautomatic segmentation approaches are highly consistent with slice-by-slice contouring by hand and can reduce the time for tumor segmentation (35,36). Future studies may employ semi-automatic segmentation to delineate the target area, thereby minimizing errors associated with manual outlining. The small sample size is also one of the limitations of this study. It may lead to insufficient robustness in the estimation of effect size and pose a certain risk of bias.
The incremental worth of the radiomics features and other clinicopathological factors in personalized EDM risk assessment is well demonstrated by the radiomics nomogram, making it a potential tool for guiding individualized treatment planning for these patients after surgery. The radiomics nomogram is a simple, non-invasive, intuitive evaluation tool for patients with e-NSCLC after radical surgery, and it may help clinical decision makers optimize treatment plans for patients and guide individualized treatment.
Conclusions
The nomogram, incorporating the radiomics signature alongside other clinicopathological factors, effectively demonstrates the incremental value in personalized risk assessment for extracranial distant metastasis. This underscores radiomics as a potential tool for guiding individualized postoperative treatment strategies in these patients. The radiomics-based nomogram model serves as a straightforward, noninvasive, and intuitive evaluation tool for patients with early-stage NSCLC following radical surgery, aiding clinical decision-makers in optimizing treatment plans and facilitating personalized therapeutic approaches.
Supplementary
The article’s supplementary files as
Acknowledgments
None.
Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective study was exempted from ethical review, and informed consent was waived owing to the use of de-identified data.
Footnotes
Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1263/rc
Funding: This work was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (No. 2024ZD0525900); Collaborative Academic Innovation Project of Shandong Cancer Hospital (No. GF002); National Natural Science Foundation of China (No. 82172866); and Department of Science & Technology of Shandong Province (No. 2021CXGC011102).
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1263/coif). The authors have no conflicts of interest to declare.
Data Sharing Statement
Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1263/dss
References
- 1. Oudkerk M, Devaraj A, Vliegenthart R, et al. European position statement on lung cancer screening. Lancet Oncol 2017;18:e754-66.. 10.1016/S1470-2045(17)30861-6 [DOI] [PubMed] [Google Scholar]
- 2. Yang D, Liu Y, Bai C, et al. Epidemiology of lung cancer and lung cancer screening programs in China and the United States. Cancer Lett 2020;468:82-7.. 10.1016/j.canlet.2019.10.009 [DOI] [PubMed] [Google Scholar]
- 3. Goldstraw P, Chansky K, Crowley J, et al. The IASLC Lung Cancer Staging Project: Proposals for Revision of the TNM Stage Groupings in the Forthcoming (Eighth) Edition of the TNM Classification for Lung Cancer. J Thorac Oncol 2016;11:39-51.. 10.1016/j.jtho.2015.09.009 [DOI] [PubMed] [Google Scholar]
- 4. Rami-Porta R, Asamura H, Travis WD, et al. Lung cancer - major changes in the American Joint Committee on Cancer eighth edition cancer staging manual. CA Cancer J Clin 2017;67:138-55.. 10.3322/caac.21390 [DOI] [PubMed] [Google Scholar]
- 5. Hung JJ, Jeng WJ, Hsu WH, et al. Prognostic factors of postrecurrence survival in completely resected stage I non-small cell lung cancer with distant metastasis. Thorax 2010;65:241-5.. 10.1136/thx.2008.110825 [DOI] [PubMed] [Google Scholar]
- 6. Choi PJ, Jeong SS, Yoon SS. Prediction and prognostic factors of post-recurrence survival in recurred patients with early-stage NSCLC who underwent complete resection. J Thorac Dis 2016;8:152-60.. 10.3978/j.issn.2072-1439.2016.01.10 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Tian Y, He Y, Li X, et al. Novel nomograms to predict lymph node metastasis and distant metastasis in resected patients with early-stage non-small cell lung cancer. Ann Palliat Med 2021;10:2548-66.. 10.21037/apm-20-1756 [DOI] [PubMed] [Google Scholar]
- 8. Choi PJ, Jeong SS, Yoon SS. Prognosis of recurrence after complete resection in early-stage non-small cell lung cancer. Korean J Thorac Cardiovasc Surg 2013;46:449-56.. 10.5090/kjtcs.2013.46.6.449 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Yang Y, Xu L, Sun L, et al. Machine learning application in personalised lung cancer recurrence and survivability prediction. Comput Struct Biotechnol J 2022;20:1811-20.. 10.1016/j.csbj.2022.03.035 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Schneider BJ, Ismaila N, Aerts J, et al. Lung Cancer Surveillance After Definitive Curative-Intent Therapy: ASCO Guideline. J Clin Oncol 2020;38:753-66.. 10.1200/JCO.19.02748 [DOI] [PubMed] [Google Scholar]
- 11. Xu X, Chen L, Xue M, et al. A global systematic analysis and statistical evaluation of prognostic prediction models for non-small cell lung cancer. The Innovation Medicine 2026;4:100195. [Google Scholar]
- 12. Zhou L, Zhang Y, Chen W, et al. Development and validation of a prognostic nomogram for early stage non-small cell lung cancer: a study based on the SEER database and a Chinese cohort. BMC Cancer 2022;22:980.. 10.1186/s12885-022-10067-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Shang X, Yu H, Lin J, et al. A Novel Nomogram including AJCC Stages Could Better Predict Survival for NSCLC Patients Who Underwent Surgery: A Large Population-Based Study. J Oncol 2020;2020:7863984.. 10.1155/2020/7863984 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Khorrami M, Bera K, Leo P, et al. Stable and discriminating radiomic predictor of recurrence in early stage non-small cell lung cancer: Multi-site study. Lung Cancer 2020;142:90-7.. 10.1016/j.lungcan.2020.02.018 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Huang Y, Liu Z, He L, et al. Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 2016;281:947-57.. 10.1148/radiol.2016152234 [DOI] [PubMed] [Google Scholar]
- 16. Fan L, Fang M, Tu W, et al. Radiomics Signature: A Biomarker for the Preoperative Distant Metastatic Prediction of Stage I Nonsmall Cell Lung Cancer. Acad Radiol 2019;26:1253-61.. 10.1016/j.acra.2018.11.004 [DOI] [PubMed] [Google Scholar]
- 17. Xu F, Li J, Xing P, et al. Heterogeneity in intracranial relapses after complete resection of lung adenocarcinoma: Distinct features of brain-only relapse versus synchronous extracranial relapse. Cancer Med 2023;12:12495-503.. 10.1002/cam4.5961 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Aerts HJ, Velazquez ER, Leijenaar RT, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5:4006.. 10.1038/ncomms5006 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Leger S, Zwanenburg A, Pilz K, et al. A comparative study of machine learning methods for time-to-event survival data for radiomics risk modelling. Sci Rep 2017;7:13206.. 10.1038/s41598-017-13448-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Harrell FE, Jr, Califf RM, Pryor DB, et al. Evaluating the yield of medical tests. JAMA 1982;247:2543-6. [PubMed] [Google Scholar]
- 21. Kattan MW. Judging new markers by their ability to improve predictive accuracy. J Natl Cancer Inst 2003;95:634-5.. 10.1093/jnci/95.9.634 [DOI] [PubMed] [Google Scholar]
- 22. Vickers AJ, Cronin AM, Elkin EB, et al. Extensions to decision curve analysis, a novel method for evaluating diagnostic tests, prediction models and molecular markers. BMC Med Inform Decis Mak 2008;8:53.. 10.1186/1472-6947-8-53 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6.. 10.1016/j.ejca.2011.11.036 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Parmar C, Rios Velazquez E, Leijenaar R, et al. Robust Radiomics feature quantification using semiautomatic volumetric segmentation. PLoS One 2014;9:e102107.. 10.1371/journal.pone.0102107 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Martin J, Ginsberg RJ, Venkatraman ES, et al. Long-term results of combined-modality therapy in resectable non-small-cell lung cancer. J Clin Oncol 2002;20:1989-95.. 10.1200/JCO.2002.08.092 [DOI] [PubMed] [Google Scholar]
- 26. Huang YQ, Liang CH, He L, et al. Development and Validation of a Radiomics Nomogram for Preoperative Prediction of Lymph Node Metastasis in Colorectal Cancer. J Clin Oncol 2016;34:2157-64.. 10.1200/JCO.2015.65.9128 [DOI] [PubMed] [Google Scholar]
- 27. Akinci D'Antonoli T, Farchione A, Lenkowicz J, et al. CT Radiomics Signature of Tumor and Peritumoral Lung Parenchyma to Predict Nonsmall Cell Lung Cancer Postsurgical Recurrence Risk. Acad Radiol 2020;27:497-507.. 10.1016/j.acra.2019.05.019 [DOI] [PubMed] [Google Scholar]
- 28. Chen A, Lu L, Pu X, et al. CT-Based Radiomics Model for Predicting Brain Metastasis in Category T1 Lung Adenocarcinoma. AJR Am J Roentgenol 2019;213:134-9.. 10.2214/AJR.18.20591 [DOI] [PubMed] [Google Scholar]
- 29. Wang L, Dong T, Xin B, et al. Integrative nomogram of CT imaging, clinical, and hematological features for survival prediction of patients with locally advanced non-small cell lung cancer. Eur Radiol 2019;29:2958-67.. 10.1007/s00330-018-5949-2 [DOI] [PubMed] [Google Scholar]
- 30. Burdett S, Rydzewska L, Tierney JF, et al. A closer look at the effects of postoperative radiotherapy by stage and nodal status: updated results of an individual participant data meta-analysis in non-small-cell lung cancer. Lung Cancer 2013;80:350-2.. 10.1016/j.lungcan.2013.02.005 [DOI] [PubMed] [Google Scholar]
- 31. Zhong WZ, Wang Q, Mao WM, et al. Gefitinib versus vinorelbine plus cisplatin as adjuvant treatment for stage II-IIIA (N1-N2) EGFR-mutant NSCLC (ADJUVANT/CTONG1104): a randomised, open-label, phase 3 study. Lancet Oncol 2018;19:139-48.. 10.1016/S1470-2045(17)30729-5 [DOI] [PubMed] [Google Scholar]
- 32. Wu YL, Tsuboi M, He J, et al. Osimertinib in Resected EGFR-Mutated Non-Small-Cell Lung Cancer. N Engl J Med 2020;383:1711-23.. 10.1056/NEJMoa2027071 [DOI] [PubMed] [Google Scholar]
- 33. Pathak R, Goldberg SB, Canavan M, et al. Association of Survival With Adjuvant Chemotherapy Among Patients With Early-Stage Non-Small Cell Lung Cancer With vs Without High-Risk Clinicopathologic Features. JAMA Oncol 2020;6:1741-50.. 10.1001/jamaoncol.2020.4232 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Bradbury P, Sivajohanathan D, Chan A, et al. Postoperative Adjuvant Systemic Therapy in Completely Resected Non-Small-Cell Lung Cancer: A Systematic Review. Clin Lung Cancer 2017;18:259-273.e8.. 10.1016/j.cllc.2016.07.002 [DOI] [PubMed] [Google Scholar]
- 35. Velazquez ER, Parmar C, Jermoumi M, et al. Volumetric CT-based segmentation of NSCLC using 3D-Slicer. Sci Rep 2013;3:3529.. 10.1038/srep03529 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Qiu Q, Duan J, Duan Z, et al. Reproducibility and non-redundancy of radiomic features extracted from arterial phase CT scans in hepatocellular carcinoma patients: impact of tumor segmentation variability. Quant Imaging Med Surg 2019;9:453-64.. 10.21037/qims.2019.03.02 [DOI] [PMC free article] [PubMed] [Google Scholar]