Predicting overall survival and prophylactic cranial irradiation benefit in small cell lung cancer patients: a multicenter cohort study Authors: Xiaomin Zheng, Kaicai Liu, Zhao Gao, Cuiping Li, Li Tong, Chang Rong, Shuai Li, Yichao Liu, Xingwang Wu Published: 2024 ABSTRACT BackgroundTo construct a CT-based radiomics nomogram, enabling the estimation of overall survival (OS) in small cell lung cancer (SCLC) patients and facilitating the identification of prophylactic cranial irradiation (PCI) beneficiaries through risk stratification using the radiomics score (RS).MethodsA retrospective recruitment of 375 patients with pathologically confirmed SCLC was conducted across three medical centers, followed by their division into different cohorts. To generate the RS, a series of analyses were performed, including Pearson correlation analysis, univariate Cox analysis, and least absolute shrinkage and selection operator (LASSO) Cox regression analysis. Subsequently, patients were stratified into either the low RS or high RS group, determined by identifying the optimal RS cutoff value. Subsequently, a radiomics nomogram was constructed using the RS, followed by assessments of its discrimination, calibration, clinical utility and reclassification. Moreover, we evaluated the potential benefits of PCI following stratification by RS.ResultsFor the internal and external validation cohorts, the radiomics nomogram (concordance index [C-index]: 0.770, 0.763) outperformed clinical nomogram (C-index: 0.625, 0.570) in predicting OS. Besides, patients with high RS had survival benefit from PCI in both the limited and extensive stage (hazard ratio [HR]: 0.304, 95% confidence interval [CI]: 0.087–1.065, P = 0.003; HR: 0.481, 95% CI: 0.270–0.860, P = 0.019, respectively), while no significant association were observed in patients with low RS.ConclusionA radiomics nomogram based on CT shows potential in predicting OS for individuals with SCLC. The RS could assist in tailoring treatment plans to identify patients likely to benefit from PCI.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12885-024-13274-7. Introduction Lung cancer remains the leading cause of cancer-related deaths worldwide, with approximately 130,180 fatalities accounting for 21% of all cancer-related deaths in the United States in 2022 [1]. Small cell lung cancer (SCLC), a type of lung cancer, constitutes approximately 15% of all lung cancers cases [2]. This neuroendocrine tumor is characterized by high aggressiveness, poor differentiation and early metastasis, resulting in an unfavorable prognosis. The Veteran Affairs Lung Group (VALG) criteria classify SCLC into limited and extensive stages, the latter comprises approximately 65% of new cases [3]. Regrettably, the prognosis for patients with SCLC is poor, with 5-year overall survival (OS) ranging from 20 to 25% for patients with limited stage, while survival is only about 2% for patients with extensive stage [4]. In 1999, a comprehensive meta-analysis demonstrated a 5.4% improvement in 3-year survival for SCLC patients receiving prophylactic cranial irradiation (PCI), especially in those with limited stage disease [5]. Due to the high risk of brain metastases in SCLC, the National Comprehensive Cancer Network (NCCN) guidelines still recommend PCI for limited stage SCLC patients after comprehensive treatment. However, brain magnetic resonance imaging (MRI) was not commonly employed in the era of the previous meta-analysis, and etoposide or irinotecan combined with platinum was not a first-line regimen in many included studies, nor was precise radiotherapy widely used, which greatly affected the initial staging and ultimate efficacy of patients, rendering the aforementioned conclusions contentious. A recent study in the United States involving 297 patients with limited stage SCLC found that PCI did not reduce in brain metastases or improve in OS [6]. Similarly, a prospective clinical trial in Japan revealed that PCI did not enhance survival, even though it reduced the rate of brain metastases in patients with extensive stage SCLC [7]. In contrast, medical practice in China continues to consider PCI following radiotherapy and chemotherapy as standard for patients with limited stage SCLC. Nonetheless, PCI adds financial strain on patients and negatively affects their cognitive function [8]. Therefore, it is of clinical significance to find out the part of patients who do not benefit from PCI and make them avoid PCI and choose observation. Initially presented by Dutch scholar Kumar in 2012 [9], radiomics involves high-throughput extraction of a wide array of features from medical images. Subsequently, machine learning is utilized to construct models, facilitating a deeper exploration, prediction, and analysis of these image-derived quantitative features. Recently, radiomics has extended its application to encompass tumor segmentation, feature extraction, and model development. This approach enables the non-invasive capture of intra-tumoral heterogeneity and the provision of physiological and pathological information associated with patient prognosis [10–12]. Within clinical application, computed tomography (CT) imaging finds widespread usage in the diagnosis and staging of SCLC while additionally serving the purpose of treatment response monitoring. Radiomics biomarkers derived from CT images have been proven to possess prognostic relevance in non-SCLC [13, 14]. However, the comparatively lower frequency of SCLC in comparison to other lung cancer subtypes has led to a scarcity of detailed studies assessing the influence of radiomics on prognosis. Moreover, there are currently no studies on the relationship between radiomics and the benefits of PCI. Therefore, our current study focused on developing and validating a CT-based radiomics nomogram to predict the OS of individuals with SCLC, and to evaluate its additional value to clinical risk factors in estimating individual OS. With this model, we further identify patients who are likely to benefit from PCI. Methods Patients and study design This retrospective study involving human participants was reviewed and approved by the ethics committees of The First Affiliated Hospital of Anhui Medical University (PJ2022-13–48), The First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, University of Science and Technology of China (2024-RE-101), and Anhui Provincial Cancer Hospital (2023-YXK-13). Informed consent was waived due to the retrospective nature. Initially, 569 patients from three medical centers were screened, and after applying the inclusion and exclusion criteria, 375 patients were enrolled. The detailed inclusion and exclusion criteria are shown in Appendix E1 and Fig. S1. Among the 375 patients, 310 were drawn from medical centers I (n = 171, January 1, 2013, to January 1, 2023) and II (n = 139, January 1, 2015, to January 1, 2023). They were then randomly assigned to a training cohort and an internal validation cohort, consisting of 217 and 93 cases, respectively, maintaining a 7:3 ratio. Meanwhile, medical center III contributed 65 patients from July 1, 2017, to January 1, 2023, forming the external validation cohort. Follow-up protocol In this study, follow-up data from institutions were obtained through outpatient and inpatient records, as well as telephone interviews. Patients were monitored every three months during the first two years post-treatment, every six months between years three and five, and annually after that. The end point of this study was OS, defined as the time between treatment initiation and the point of either mortality or the latest follow-up occurrence. The median follow-up time for all patients with OS was 13 months, varying from a minimum of 3 months post-treatment to a maximum of 63 months. Treatment protocol Chemotherapy serves as the primary treatment protocol. First-line chemotherapy drugs are mainly etoposide combined with cisplatin/carboplatin, administered over a total of 4 to 6 cycles, with each cycle lasting approximately 21 days. All chest radiotherapy methods used are either intensity-modulated radiotherapy or three-dimensional conformal radiotherapy. In contrast, PCI is a widely used technique that employs two opposing fields for penetration. Chest radiotherapy is administered with a total dose ranging from 50.4 to 66 Gy, provided at a daily dose of 1.8 to 2.0 Gy, given five times weekly. As for PCI, it is commonly prescribed at a dose of 30 Gy over ten fractions. CT scanning and image acquisition Patients enrolled across the three medical centers underwent CT scans with comparable setups and parameters, although different CT scanners were used. Detailed information about the CT image acquisition settings is provided in Table S1 and Appendix E2. Tumor segmentation and feature extraction The workflow of this study is presented in Fig. 1. Two radiologists, one with 5 years of experience (reader 1,) and the other with 8 years (reader 2, K.L.), independently carried out 3D tumor segmentation using ITK-SNAP software (version 3.8). Layer by layer, the region of interest (ROI) was carefully drawn along the primary tumor’s border within the enhanced CT images captured during the arterial phase. These images were displayed using the mediastinal window settings (level, 40 HU; width, 400 HU). This process extended until the entire tumor area was covered, leading to the creation of 3D tumor segmentation. It’s worth noting the careful exclusion of blood vessels and airs from the ROI. Appendix E3 illustrates the procedures for assessing inter-/intra-observer reproducibility. Radiomics features were extracted using PyRadiomics (version 3.0.1) in Python (version 3.6.0). Comprehensive details regarding feature extraction can be found in Appendix E4.Fig. 1The radiomics flow chart of the study. ROI, region of interest; GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run length matrix; GLSZM, gray-level size zone matrix; NGTDM, neighboring gray tone difference matrix; ICC, inter-/intraclass correlation coefficient; LASSO, least absolute shrinkage and selection operator; OS, overall survival; ROC, receiver operating characteristic; AUC, area under the curve The radiomics flow chart of the study. ROI, region of interest; GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run length matrix; GLSZM, gray-level size zone matrix; NGTDM, neighboring gray tone difference matrix; ICC, inter-/intraclass correlation coefficient; LASSO, least absolute shrinkage and selection operator; OS, overall survival; ROC, receiver operating characteristic; AUC, area under the curve Feature selection and radiomics score (RS) calculation Images from the training cohort underwent feature selection. Firstly, reproducibility was assessed by calculating inter-/intraclass correlation coefficients (ICCs), and features with ICC > 0.75 were retained. Secondly, redundant features were eliminated using Pearson correlation analysis, where pairs of features with an absolute Pearson correlation coefficient exceeding 0.8 were evaluated, discarding features with a higher mean absolute correlation. Subsequently, univariate Cox analysis was employed to identify features with a significant association with OS (P < 0.05). Lastly, the least absolute shrinkage and selection operator (LASSO) Cox regression analysis was utilized to pinpoint the most valuable prognostic features, and a RS was computed for each patient by linearly combining the selected features weighted by their corresponding coefficients. Radiomics nomogram construction and validation Patients were stratified into low and high RS groups based on the RS in the training cohort, which was determined by the optimal cutoff value identified with the X-tile program (version 3.6.1). Kaplan–Meier survival analysis was performed to investigate the association between the RS and OS for each cohort. Afterwards, a radiomics nomogram was constructed based on the RS. To assess the accuracy of the radiomics nomogram in predicting patient survival, time-dependent receiver operating characteristic (ROC) curves were plotted for the 1-year, 2-year, and 3-year OS in all cohorts, and the corresponding area under the curves (AUCs) were calculated. To further evaluate the effectiveness of the radiomics nomogram, we employed Harrell’s concordance index (C-index), calibration curves, and conducted decision curve analysis (DCA). Additionally, we quantified the increased prognostic enhancement offered by the radiomics nomogram through net reclassification improvement (NRI) [15] and integrated discrimination improvement [IDI] [16]. To demonstrate the enhanced utility of the radiomics nomogram, we developed a clinical nomogram and a combined clinical-radiomics nomogram. These models were established in the training cohort and subsequently assessed in validation cohorts (details are shown in Appendix E5). RS and PCI Radiomics features have been shown to correlate with the effectiveness of anti-tumor therapies across different cancer types [17, 18]. Thus, we sought to explore the possible relationship between the RS and the efficacy of PCI in all patients. Additionally, we performed stratified analyses according to clinical stage and RS to identify subgroups of patients who might benefit from PCI. Moreover, interaction tests were also performed to examine the relationships between the RS, clinical stage, and PCI. Statistical analysis Continuous variables were expressed as mean ± standard deviation for those with a normal distribution, whereas median (interquartile range [IQR]) was applied to variables that did not follow a normal distribution. Group comparisons were conducted using the Mann–Whitney U test, t test, or Kruskal–Wallis H test. Categorical variables were analyzed between groups using either the Chi-square test or Fisher’s exact test. Variables with a P value less than 0.05 in the univariate Cox analysis were incorporated into the multivariate Cox regression. Model performance was evaluated through Harrell’s C-index, time-dependent ROC curves, calibration plots, DCA, as well as NRI and IDI. All statistical analyses were performed using R software (version 3.5.3, http://www.r-project.org) and SPSS (IBM version 25.0). The used source codes are available at GitHub (https://github.com/zxm081921/SCLC-Radiomics). Significance was determined using a two-sided P-value threshold of < 0.05. Results Patient characteristics In the study, Table 1 presents the clinical factors of the 375 patients who participated. The training cohort had a median follow-up duration for OS of 13.0 months, with an IQR of 10.0 to 22.0 months. Likewise, the internal validation cohort exhibited a median follow-up period of 15.0 months, with an IQR of 9.0 to 29.5 months, and the external validation cohort had a median follow-up of 12.0 months, with an IQR of 9.0 to 28.0 months. There were no significant differences in clinical variables across the three cohorts (P = 0.104–0.902). Table 1Clinical characteristics of patients with small cell lung cancer in the training and validation cohortsVariablesTraining cohort(n = 217)Internal validation cohort(n = 93)External validation cohort(n = 65)P valueAge, no. (%)0.104 < 65 years1244628 ≥ 65 years934737Gender, no. (%)0.533 Female113 (52.1)42 (45.2)33 (50.8) Male104 (47.9)51 (54.8)32 (49.2)Smoking history, no. (%)0.203 Nonsmoker115 (53.0)51 (54.8)27 (41.5) Smoker102 (47.0)42 (45.2)38 (58.5)Stage, no. (%)0.819 Limited stage58 (26.7)23 (24.7)15 (23.1) Extensive stage159 (73.3)70 (75.3)50 (76.9)Chemotherapy cycle, no. (%)0.803 ≤ 482 (37.8)38 (40.9)27 (41.5) > 4135 (62.2)55 (59.1)38 (58.5)Radiotherapy, no. (%)0.823 Yes81 (37.3)35 (37.6)27 (41.5) No136 (62.7)58 (62.4)38 (58.5)PCI, no. (%)0.902 Yes42 (19.4)17 (18.3)11 (16.9) No175 (80.6)76 (81.7)54 (83.1)Tumor size (cm, mean ± SD)4.9 ± 2.64.7 ± 2.35.3 ± 2.60.323CEA, no. (%)0.122 ≤ 5 ng/ml1235045 > 5 ng/ml944320NSE, no. (%)0.188 ≤ 16.3 ng/ml15109 > 16.3 ng/ml2028356SD Standard deviation, PCI Prophylactic cranial irradiation, CEA Carcinoma embryonic antigen, NSE Neuron specific enolase, IQR Interquartile range Table 1Clinical characteristics of patients with small cell lung cancer in the training and validation cohortsVariablesTraining cohort(n = 217)Internal validation cohort(n = 93)External validation cohort(n = 65)P valueAge, no. (%)0.104 < 65 years1244628 ≥ 65 years934737Gender, no. (%)0.533 Female113 (52.1)42 (45.2)33 (50.8) Male104 (47.9)51 (54.8)32 (49.2)Smoking history, no. (%)0.203 Nonsmoker115 (53.0)51 (54.8)27 (41.5) Smoker102 (47.0)42 (45.2)38 (58.5)Stage, no. (%)0.819 Limited stage58 (26.7)23 (24.7)15 (23.1) Extensive stage159 (73.3)70 (75.3)50 (76.9)Chemotherapy cycle, no. (%)0.803 ≤ 482 (37.8)38 (40.9)27 (41.5) > 4135 (62.2)55 (59.1)38 (58.5)Radiotherapy, no. (%)0.823 Yes81 (37.3)35 (37.6)27 (41.5) No136 (62.7)58 (62.4)38 (58.5)PCI, no. (%)0.902 Yes42 (19.4)17 (18.3)11 (16.9) No175 (80.6)76 (81.7)54 (83.1)Tumor size (cm, mean ± SD)4.9 ± 2.64.7 ± 2.35.3 ± 2.60.323CEA, no. (%)0.122 ≤ 5 ng/ml1235045 > 5 ng/ml944320NSE, no. (%)0.188 ≤ 16.3 ng/ml15109 > 16.3 ng/ml2028356SD Standard deviation, PCI Prophylactic cranial irradiation, CEA Carcinoma embryonic antigen, NSE Neuron specific enolase, IQR Interquartile range Clinical characteristics of patients with small cell lung cancer in the training and validation cohorts SD Standard deviation, PCI Prophylactic cranial irradiation, CEA Carcinoma embryonic antigen, NSE Neuron specific enolase, IQR Interquartile range Feature selection and RS calculation Using CT images from the arterial phase with the mediastinal window, we extracted a total of 1688 radiomics features. In the training cohort, we calculated intra-observer and inter-observer ICCs for the radiomic features, finding ranges of 0.334 to 0.941 and 0.415 to 0.938, respectively. We retained 1357 radiomics features that had ICCs greater than 0.75 for both inter- and intra-observer reliability. After performing Pearson correlation analysis to eliminate redundant features, 142 radiomics features remained in the dataset. Subsequently, following univariate Cox analysis, it was found that 18 features had a significant association with patients’ OS (P < 0.05). Finally, the LASSO Cox regression algorithm was used to extract 9 radiomics features successfully (Fig. 2A-B). The formula for calculating RS is shown in the Appendix E6. Figure 2C illustrates a histogram showing the presentation of the selected features along with their corresponding coefficients contributing to the RS calculation.Fig. 2Radiomics features were selected by the least absolute shrinkage and selection operator (LASSO) Cox regression model. A LASSO coefficient profiles of 18 radiomics features; (B) Tuning parameter (λ) selection in the LASSO model via tenfold cross-validation based on minimum criterion. Vertical lines were drawn at the optimal values. The optimal λ (λ = 0.085) resulted in 9 radiomics features with nonzero coefficients; (C) Histogram shows the role of individual features that contribute to the radiomics score Radiomics features were selected by the least absolute shrinkage and selection operator (LASSO) Cox regression model. A LASSO coefficient profiles of 18 radiomics features; (B) Tuning parameter (λ) selection in the LASSO model via tenfold cross-validation based on minimum criterion. Vertical lines were drawn at the optimal values. The optimal λ (λ = 0.085) resulted in 9 radiomics features with nonzero coefficients; (C) Histogram shows the role of individual features that contribute to the radiomics score Survival risk stratification based on the RS According to the optimal RS cutoff of 0.14 determined through X-tile analysis in the training cohort, patients were divided into two categories: those with a low RS (RS < 0.14) and those with a high RS (RS ≥ 0.14). The distribution of RS revealed that patients with low RS experienced more favorable survival outcomes, while those with high RS faced poorer OS, as depicted in Fig. 3A. Survival analysis demonstrated a substantial disparity in prognosis between patients with high RS and those with low RS across all cohorts. In the training cohort (hazard ratio [HR] = 4.025; 95% confidence interval [CI]: 2.563–6.320; P < 0.001; Fig. 3B), the internal validation cohort (HR = 2.836; 95% CI: 1.560–5.155; P < 0.001; Fig. 3C), and the external validation cohort (HR = 3.579; 95% CI: 1.548–8.272; P < 0.001; Fig. 3D), patients with high RS exhibited significantly worse prognostic outcomes when compared to their low RS counterparts. Moreover, subgroup analysis indicated that patients with a high RS had significantly shorter OS compared to those with a low RS, regardless of whether they were in the limited or extensive stage (Fig. S2). These findings validate the strong prognostic accuracy associated with the RS.Fig. 3Survival evaluation based on the RS. A Value distribution of the RS; (B) Kaplan–Meier survival curves of patients at different risks stratified by the RS in the training cohort, internal validation cohort (C) and external validation cohort (D). RS, radiomics score Survival evaluation based on the RS. A Value distribution of the RS; (B) Kaplan–Meier survival curves of patients at different risks stratified by the RS in the training cohort, internal validation cohort (C) and external validation cohort (D). RS, radiomics score Incremental prognostic value of the radiomics nomogram Table 2 presents the findings from both univariable and multivariable Cox regression analyses regarding predictors of OS in the training cohort. The results indicate that stage and chemotherapy cycles are clinically independent prognostic variables (stage: HR: 2.065, 95% CI: 1.450–2.939, P < 0.001; chemotherapy cycle: HR: 0.561, 95% CI: 0.411–0.767, P < 0.001). In addition, stage, chemotherapy cycles, and RS are clinical-radiomics independent prognostic variables (stage: HR: 1.731, 95% CI: 1.213–2.469, P = 0.002; chemotherapy cycle: HR: 0.654, 95% CI: 0.478–0.895, P = 0.008; RS: HR: 7.978, 95% CI: 5.318–11.969, P < 0.001). Table 2Univariate and multivariable relationships between overall survival, radiomics score, and clinical characteristicsVariablesUnivariate Cox RegressionMultivariable Cox Regression ClinicalClinical-radiomicsHR (95% CI)P valueHR (95% CI)P valueHR (95% CI)P valueAge (< 65 years vs ≥ 65 years)0.942 (0.697–1.275)0.700Gender (Female vs Male)1.122 (0.770–1.634)0.550Smoking history (Nonsmoker vs Smoker)0.840 (0.621–1.135)0.255Stage (Limited vs Extensive)2.014 (1.414–2.868) < 0.0012.065 (1.450–2.939) < 0.0011.731 (1.213–2.469)0.002Chemotherapy cycle ( ≤ 4 vs > 4)0.579 (0.424–0.791)0.0010.561 (0.411–0.767) < 0.0010.654 (0.478–0.895)0.008Radiotherapy (Yes vs No)0.801 (0.585–1.096)0.165PCI (Yes vs No)0.721 (0.535–0.971)0.0311.060 (0.682, 1.648)0.7961.054 (0.681, 1.630)0.814Tumor size (cm)0.989 (0.928–1.055)0.745CEA (≤ 5 ng/ml vs > 5 ng/ml)0.972 (0.644–1.181)0.375NSE (≤ 16.3 ng/ml vs > 16.3 ng/ml)0.893 (0.507–1.574)0.696Radiomics score8.298 (5.608–12.278) < 0.0017.978 (5.318–11.969) < 0.001HR Hazard ratio, CI Confidence interval, PCI Prophylactic cranial irradiation, CEA Carcinoma embryonic antigen, NSE Neuron specific enolase Table 2Univariate and multivariable relationships between overall survival, radiomics score, and clinical characteristicsVariablesUnivariate Cox RegressionMultivariable Cox Regression ClinicalClinical-radiomicsHR (95% CI)P valueHR (95% CI)P valueHR (95% CI)P valueAge (< 65 years vs ≥ 65 years)0.942 (0.697–1.275)0.700Gender (Female vs Male)1.122 (0.770–1.634)0.550Smoking history (Nonsmoker vs Smoker)0.840 (0.621–1.135)0.255Stage (Limited vs Extensive)2.014 (1.414–2.868) < 0.0012.065 (1.450–2.939) < 0.0011.731 (1.213–2.469)0.002Chemotherapy cycle ( ≤ 4 vs > 4)0.579 (0.424–0.791)0.0010.561 (0.411–0.767) < 0.0010.654 (0.478–0.895)0.008Radiotherapy (Yes vs No)0.801 (0.585–1.096)0.165PCI (Yes vs No)0.721 (0.535–0.971)0.0311.060 (0.682, 1.648)0.7961.054 (0.681, 1.630)0.814Tumor size (cm)0.989 (0.928–1.055)0.745CEA (≤ 5 ng/ml vs > 5 ng/ml)0.972 (0.644–1.181)0.375NSE (≤ 16.3 ng/ml vs > 16.3 ng/ml)0.893 (0.507–1.574)0.696Radiomics score8.298 (5.608–12.278) < 0.0017.978 (5.318–11.969) < 0.001HR Hazard ratio, CI Confidence interval, PCI Prophylactic cranial irradiation, CEA Carcinoma embryonic antigen, NSE Neuron specific enolase Univariate and multivariable relationships between overall survival, radiomics score, and clinical characteristics HR Hazard ratio, CI Confidence interval, PCI Prophylactic cranial irradiation, CEA Carcinoma embryonic antigen, NSE Neuron specific enolase In Fig. S3A, Fig. 4A and Fig. S4A, we present the clinical nomogram, radiomics nomogram and clinical-radiomics nomogram, respectively, both constructed using the training cohort. The radiomics nomogram (C-index: 0.770; 1-, 2-, 3-year AUC: 0.850, 0.867, 0.857) and the clinical-radiomics nomogram (C-index: 0.793; 1-, 2-, 3-year AUC: 0.852, 0.891, 0.862) exhibited significantly better predictive capabilities for survival risks among SCLC patients compared to the clinical nomogram alone (C-index: 0.625; 1-, 2-, 3-year AUC: 0.642, 0.671, 0.654), a finding that was corroborated in the external validation cohort (Table 3, Fig. 4A, Fig. S3B and Fig. S4B). Additionally, time-dependent ROC analysis at various time points confirmed the radiomics nomogram’s additional prognostic performance (Fig. 5A).Fig. 4Use of the constructed radiomics nomogram to estimate OS for SCLC, along with the assessment of the time-dependent ROC curves and calibration curves. A The radiomics nomogram for estimating OS; (B) AUCs at 1-year, 2-year, and 3-year OS were calculated to assess the prognostic accuracy of the clinical-radiomics nomogram within the training cohort (left), internal validation cohort (middle) and external validation cohort (right); (C) Plot shows calibration of the clinical-radiomics nomogram in terms of agreement between predicted and observed 1-year, 2-year, and 3-year OS outcomes in the training cohort (left), internal validation cohort (middle) and external validation cohort (right). OS, overall survival; SCLC, small cell lung cancer; ROC, receiver operating characteristic; AUC, area under the curveTable 3The performances of the different models in the training and validation cohortsCohortC-index (95% CI)Clinical nomogramRadiomics signatureClinical-radiomics nomogramTraining cohort0.628 (0.584, 0.672)0.787 (0.752, 0.823)0.803 (0.768, 0.837)Internal validation cohort0.625 (0.555, 0.695)0.770 (0.706, 0.835)0.793 (0.732, 0.853)External validation cohort0.570 (0.467, 0.674)0.763 (0.659, 0.867)0.770 (0.665, 0.875)C-index Concordance index, CI Confidence intervalFig. 5Prognostic performance and decision curve analysis of the clinical nomogram, radiomics nomogram and clinical-radiomics nomogram. A Time-dependent areas under the ROC curve of the clinical nomogram, radiomics nomogram and clinical-radiomics nomogram in the training cohort (left), internal validation cohort (middle) and external validation cohort (right). B The decision curve analysis of the clinical nomogram, radiomics nomogram and clinical-radiomics nomogram in the training cohort (left), internal validation cohort (middle) and external validation cohort (right). ROC, receiver operating characteristic; AUC, area under the curve Use of the constructed radiomics nomogram to estimate OS for SCLC, along with the assessment of the time-dependent ROC curves and calibration curves. A The radiomics nomogram for estimating OS; (B) AUCs at 1-year, 2-year, and 3-year OS were calculated to assess the prognostic accuracy of the clinical-radiomics nomogram within the training cohort (left), internal validation cohort (middle) and external validation cohort (right); (C) Plot shows calibration of the clinical-radiomics nomogram in terms of agreement between predicted and observed 1-year, 2-year, and 3-year OS outcomes in the training cohort (left), internal validation cohort (middle) and external validation cohort (right). OS, overall survival; SCLC, small cell lung cancer; ROC, receiver operating characteristic; AUC, area under the curve Table 3The performances of the different models in the training and validation cohortsCohortC-index (95% CI)Clinical nomogramRadiomics signatureClinical-radiomics nomogramTraining cohort0.628 (0.584, 0.672)0.787 (0.752, 0.823)0.803 (0.768, 0.837)Internal validation cohort0.625 (0.555, 0.695)0.770 (0.706, 0.835)0.793 (0.732, 0.853)External validation cohort0.570 (0.467, 0.674)0.763 (0.659, 0.867)0.770 (0.665, 0.875)C-index Concordance index, CI Confidence interval The performances of the different models in the training and validation cohorts C-index Concordance index, CI Confidence interval Prognostic performance and decision curve analysis of the clinical nomogram, radiomics nomogram and clinical-radiomics nomogram. A Time-dependent areas under the ROC curve of the clinical nomogram, radiomics nomogram and clinical-radiomics nomogram in the training cohort (left), internal validation cohort (middle) and external validation cohort (right). B The decision curve analysis of the clinical nomogram, radiomics nomogram and clinical-radiomics nomogram in the training cohort (left), internal validation cohort (middle) and external validation cohort (right). ROC, receiver operating characteristic; AUC, area under the curve The calibration plots showed good agreement between the estimations and the clinical outcomes for the radiomics nomogram and clinical-radiomics nomogram at 1-, 2-, and 3-year, in contrast to the clinical nomogram (Fig. 4B, Fig. S3C and Fig. S4C). Moreover, as shown in Fig. 5B, the DCA highlighted that both the radiomics nomogram and the clinical-radiomics nomogram provided superior overall net benefits across a spectrum of reasonable threshold probabilities when compared to the clinical nomogram. Furthermore, when the radiomics nomogram was integrated into the clinical nomogram, it resulted in a notable increase in the NRI by 32.4% (95% CI: 0.106–0.571, P < 0.05) and an IDI of 18.8% (95% CI: 0.051–0.292, P < 0.05) in the internal validation cohorts. Similar findings were replicated in the external validation cohort, as illustrated in Fig. S5. Nevertheless, no significant differences were observed in the NRI and IDI between the radiomics nomogram and clinical-radiomics nomogram in either the internal or external validation cohort (all P > 0.05). PCI benefit based on RS risk stratification In order to identify patients who may benefit from PCI, interaction tests were conducted among RS, clinical stage, and PCI efficacy (Table 4). The interaction test indicated that patients with a high RS had a better PCI benefit (HR: 0.465, 95% CI: 0.308–0.702, P < 0.001; P = 0.008 for interaction) compared to those with a low RS. Kaplan–Meier survival curves for low and high RS groups are shown in Fig. 6A. Furthermore, the interaction tests for RS and PCI efficacy in the limited and extensive stage patients were also performed. Subgroup analysis based on stage indicated that patients with high RS gained benefits from PCI in both stages (HR: 0.304, 95% CI: 0.087–1.065, P = 0.003; P = 0.001 for interaction in limited stage, and HR: 0.481, 95% CI: 0.270–0.860, P = 0.019; P = 0.017 for interaction in extensive stage). In contrast, no significant correlation was observed in the low RS group (Fig. 6B, C). Table 4Hazard ratios from PCI for associations of the overall survival with RS in patients with small cell lung cancerNo PCIPCIOverall survivalNo PCI vs PCI, HR (95% CI)PP value for interactionAll (n = 375)0.008 Low RS191460.796 (0.565, 1.121)0.206 High RS114240.465 (0.308, 0.702) < 0.001Limited stage (n = 96)0.001 Low RS33391.525 (0.895, 2.598)0.085 High RS10140.304 (0.087, 1.065)0.003Extensive stage (n = 279)0.017 Low RS15871.007 (0.469, 2.159)0.986 High RS104100.481 (0.270, 0.860)0.019PCI Prophylactic cranial irradiation, RS Radiomics score, HR Hazard ratio, CI Confidence intervalFig. 6PCI benefits based on overall survival according to RS and stage. Kaplan–Meier survival curves are shown for patients according to their use of PCI. All patients (A), limited stage patients (B), and extensive stage (C) patients were stratified according to low RS (left) and high RS (right). PCI, prophylactic cranial irradiation; RS, radiomics score Table 4Hazard ratios from PCI for associations of the overall survival with RS in patients with small cell lung cancerNo PCIPCIOverall survivalNo PCI vs PCI, HR (95% CI)PP value for interactionAll (n = 375)0.008 Low RS191460.796 (0.565, 1.121)0.206 High RS114240.465 (0.308, 0.702) < 0.001Limited stage (n = 96)0.001 Low RS33391.525 (0.895, 2.598)0.085 High RS10140.304 (0.087, 1.065)0.003Extensive stage (n = 279)0.017 Low RS15871.007 (0.469, 2.159)0.986 High RS104100.481 (0.270, 0.860)0.019PCI Prophylactic cranial irradiation, RS Radiomics score, HR Hazard ratio, CI Confidence interval Hazard ratios from PCI for associations of the overall survival with RS in patients with small cell lung cancer PCI Prophylactic cranial irradiation, RS Radiomics score, HR Hazard ratio, CI Confidence interval PCI benefits based on overall survival according to RS and stage. Kaplan–Meier survival curves are shown for patients according to their use of PCI. All patients (A), limited stage patients (B), and extensive stage (C) patients were stratified according to low RS (left) and high RS (right). PCI, prophylactic cranial irradiation; RS, radiomics score Discussion In this study we constructed radiomics nomogram with 9 CT-based radiomics features linked to OS in patients with SCLC and assessed its additional value over clinical risk factors for personalized OS estimation. Furthermore, we investigated the association between the RS and PCI, and stratified patients based on their RS to help identifying those who may benefit from PCI. Radiomics, a field advancing swiftly, shows significant potential. It illustrates intra-tumor heterogeneity and pathophysiological insights by extracting extensive quantitative features from images and constructing models through a variety of statistical techniques [19]. Radiomics provides relevant data for the identification and grading of tumors, the assessment of treatment efficacy and prognosis, and the customization of individualized treatment strategies [20, 21]. The prognosis for patients with SCLC shows considerable variability, even when they find themselves at identical stages and are subjected to uniform therapeutic regimens. This highlights the need for accurate prognosis evaluation to facilitate the execution of individualized treatment strategies. While earlier research has demonstrated the significance of radiomics features in prognostic prediction of non-SCLC [13, 14, 22], its potential in predicting the survival outcomes for SCLC patients has remained relatively unexplored. Jain et al. [23] indicated that radiomics features obtained from within and around tumors hold the capability to predict OS of SCLC patients. Conversely, in a separate investigation, Gkika et al. [24] communicated that they failed to identify any radiomics features significantly linked to the OS of SCLC patients. In view of the small sample size of the above two studies, only 153 and 98 cases respectively, the results lack reliability. Our analysis presented additional evidence from a multicenter study involving 375 patients with SCLC that CT-based radiomics features could independently predict OS. The utilization of 9 radiomics features for constructing the RS effectively segregated these patients into low and high RS groups, with Kaplan–Meier curves clearly showed significant differences in OS between different risk stratified groups. This methodology could potentially enable clinicians to choose personalized treatments of heightened efficacy for patients facing varying degrees of risk. Moreover, upon the stratification of patients according to their clinical stage, significant differences in OS between the low RS and high RS groups within both the limited and extensive stages. These findings imply a notable diversity in survival outcomes among patients categorized by RS, even when they share the same clinical stage. This finding is consistent with a study by Wang et al. [14], which demonstrated in non-SCLC that subgroup analysis by clinical and pathological stage revealed patients with high RS had significantly shorter recurrence-free survival than those with low RS, even within the same stage. A nomogram is constructed using multi-factor regression analysis, incorporating various predictive markers to estimate specific clinical outcomes or event probabilities. It demonstrates notable precision in prognosticating certain malignant tumor outcomes [25, 26]. Our investigation demonstrates that the radiomics nomogram outperforms the clinical nomogram in predictive efficacy, as evidenced by measures such as C-index, AUC, calibration curve, DCA curve, NRI, and IDI. Nevertheless, this study yielded that the clinical-radiomics nomogram held no notable superiority over the radiomics nomogram. This observation further corroborated the limitations of clinical variables, accentuating the distinct benefit of the radiomics nomogram when it comes to prognosticating OS in SCLC. This could potentially result from the limitation of clinical factors in reflecting tumor heterogeneity, unable to provide a comprehensive reflection of the overall tumor metabolism. Radiomics based on CT images can provide a comprehensive and quantitative depiction of tumors’ biological characteristics. Moreover, it adeptly elucidates the nuanced aspects present within images, thereby facilitating an exploration of potential pathophysiological information [27]. This capability might elucidate the improved efficacy in prognostication. A previous large-scale study established a clinical OS prognostic nomogram based on 24,680 patients with SCLC, and validated in an independent testing group (9,700 SCLC patients) with a C-index of 0.722 and an integrated AUC of 0.79 [28]. Given the large variation in sample size, the clinical nomogram of our study (C-index: 0.625, 0.570, AUC: 0.576–0.692 in validation cohorts) is significantly inferior to the above study, but the radiomics nomogram (C-index: 0.770, 0.763, AUC: 0.766–0.893 in validation cohorts) seems to be superior. Brain metastases are one of the most common sites of metastasis in patients with SCLC, with approximately 10% of patients developing brain metastases at first diagnosis [29]. As a preventive measure for SCLC patients at risk of brain metastases, whole-brain radiotherapy, known as PCI, is employed. A clinical study conducted by the European Organization for Research and Treatment of Cancer (EORTC) demonstrated that PCI enhances OS in patients with extensive stage SCLC [30]. Consequently, international guidelines recommend PCI for SCLC patients who have achieved complete or partial responses after chemoradiotherapy. Nevertheless, the incorporation of brain MRI in clinical practice, multiple studies have suggested that PCI might not be advantageous for all individuals [6, 7]. For instance, Pezzi et al. [6] reported that PCI did not correlate with improved OS in limited-stage SCLC (P = 0.32). Additionally, a phase 3 trial conducted involving 47 institutions in Japan found that PCI did not result in an extended OS for patients with extensive stage SCLC (P = 0.094) [7]. To further evaluate the benefits of PCI, our study was the first to introduce a risk stratification method using RS to identify SCLC patients who could potentially benefit from PCI. The most significant finding of our study was that patients with high RS, experienced significant benefits from PCI, whether in the limited or extensive stage (P = 0.003, 0.019, respectively). In contrast, those with low RS did not exhibit any clear benefit from PCI (P = 0.085, 0.986, respectively). These results suggest that, according to the RS-based stratification, PCI may be unnecessary for patients with a low RS. Prior researches have established that PCI is linked to significant neurotoxicity, leading to impairments in brain structure and cognitive function, which in turn impacts patients’ quality of life [8, 31]. This may explain why patients with low RS do not experience benefits from PCI. Additionally, a meta-analysis revealed that stage pT1–2N0M0 patients had a risk of brain metastases was just 12% even in the absence of PCI [32]. As a result, patients with low RS are less likely to develop brain metastases. In contrast, those with high RS face a worse prognosis, and PCI may provide survival benefits for them. This study has several limitations that should be noted. Firstly, owing to its retrospective nature, potential biases in the selection of the sample size might have occurred, and the verification of clinical application lacked multi-center prospective studies. Future studies will be conducted to analyze the survival benefits of different stage for SCLC based on radiomics, so as to achieve accurate personalized treatment and optimize treatment plans. Secondly, the intricacy, time consumption, and labor intensiveness associated with manual segmentation are noteworthy. Consequently, the adoption of more dependable and time-efficient alternatives, such as automated segmentation should be contemplated [33]. Lastly, the CT image slice thickness measured 5 mm, a factor that might compromise the precision of feature extraction in comparison to the thin-slice CT. In summary, a CT-based radiomics nomogram serves as an effective tool for predicting OS in patients with SCLC, offering additional value to clinical risk factors in estimating individual OS. Furthermore, risk stratification of RS could assist in determining which patients are more likely to benefit from PCI. Supplementary Information Supplementary Material 1. Supplementary Material 1.