CT-based radiomics nomogram for preoperative prediction of Ki-67 in lung neuroendocrine neoplasms: a multicenter study

Abstract

Objective

Lung neuroendocrine neoplasms (L-NENs) are increasingly recognized, yet reliable preoperative assessment of the Ki-67 proliferation index remains invasive and subject to sampling variability. We aimed to develop and validate a clinical-radiomics nomogram that uses routine chest CT to estimate Ki-67 status in patients with L-NENs.

Methods

In this retrospective multicenter study, 199 patients with histologically confirmed L-NENs from four hospitals between January 2014 and April 2024 were enrolled, all of whom underwent preoperative dual-phase contrast-enhanced CT. Following manual 3D tumor segmentation, a total of 1,874 radiomics features were extracted from fused non-contrast and arterial/venous phase images. Feature selection was performed using Pearson correlation analysis (removing redundant features with correlation coefficients > 0.8), followed by further variable compression via LASSO regression to identify discriminative radiomics features. Based on the selected features, five classification models were constructed, and the best-performing one was combined with clinical predictors identified through univariate and multivariate analyses to develop a radiomics-based nomogram. The model’s discriminative ability, calibration, and clinical utility were evaluated in the training set (n = 116), internal test set (n = 50), and external validation set (n = 33) using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA), respectively.

Results

The LR-based radiomics model demonstrated high discriminatory ability, achieving AUCs of 0.912 (95% CI: 0.858–0.965) in the training set and 0.943 (0.887–0.999) in the testing set, significantly outperforming other models. Consequently, it was combined with independent clinical predictors—largest tumor diameter, smoking history, and age—to build a nomogram. The final combined model exhibited excellent performance across all datasets, with AUCs of 0.958 (0.925–0.990) in training, 0.930 (0.865–0.995) in testing, and 0.911 (0.867–0.955) in external validation, accompanied by good calibration and a superior net benefit on decision curve analysis.

Conclusion

The CT-based clinical-radiomics nomogram provides an accurate, non-invasive tool for pre-operative Ki-67 estimation in L-NENs, potentially guiding treatment decisions. Prospective, larger-scale validation is warranted.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-025-02132-y.

Background

Lung neuroendocrine neoplasms (L-NENs), which arise from neuroendocrine and peptidergic cells in the bronchial mucosa and submucosal glands, account for approximately 20% of lung malignancies [1]. The prevalence of NEN is steadily increasing, likely due to improved diagnostic techniques and increased clinical awareness [2].

The current World Health Organization classification categorizes L-NENs based on tumor necrosis and mitotic index, supplemented by immunohistochemistry (IHC) staining for neuroendocrine markers [3]. Typical carcinoid and atypical carcinoid are classified as well-differentiated neuroendocrine tumors (NET), while large cell neuroendocrine carcinoma (LCNEC) and small cell lung carcinoma (SCLC) are categorized as poorly differentiated neuroendocrine carcinomas (NEC). Significant crushing artifacts are frequently observed in tissue fragments acquired via preoperative biopsy, which can hinder mitotic count assessment and complicate morphological interpretation. In such cases, the Ki-67 proliferation index (PI) is highly valuable [4].

The WHO lists the Ki-67 antigen, a well-known cell proliferation marker, as a non-mandatory but desirable criterion for L-NENs, with ≤ 5%, < 30%, and > 30% Ki-67-positive cells indicating typical carcinoids (TC), atypical carcinoids (AC), and neuroendocrine carcinoma, respectively [3]. Although not formally incorporated into the staging system, Ki-67 IHC testing of biopsy samples is vital for differentiating NET from NEC and preventing misdiagnosis [4, 5].

The clinical management of L-NENs varies significantly based on tumor grade and Ki-67 expression level. Radical surgery remains optimal for TC and early-stage atypical AC with Ki-67 < 30%. LCNEC and SCLC, characterized by high Ki-67 levels and high-grade malignancy, exhibit significant rates of local and systemic metastases, with treatment primarily involving a combination of radiotherapy and chemotherapy [6]. Furthermore, patients with TC with increased Ki-67 expression show significantly reduced survival [7].

Preoperative Ki-67 expression is primarily assessed through IHC testing, which involves obtaining tissue samples by puncture and evaluating them by routine visual observation by a pathologist [8–10]. Owing to tumor heterogeneity and the relatively small sample size, Ki-67 evaluation based on biopsy samples may not accurately represent the entire tumor. Furthermore, biopsy sampling is invasive and non-repeatable. Additionally, in clinical scenarios where core needle biopsy is contraindicated or not technically feasible, Ki-67 assessment cannot be performed. Therefore, there is a pressing need for reliable non-invasive techniques for accurate prediction of Ki-67 status in L-NENs.

Radiomics provides comprehensive and quantitative tumor measurements to facilitate non-invasive, comprehensive analysis of tumor phenotypes. In oncology, radiomics is increasingly employed for diagnosis, treatment planning, and prognosis evaluation [11]. Radiomics can help improve tumor behavior and patient treatment management understanding by extracting high-throughput data from medical images, leading to more personalized and precise medical care [12–15].

However, limited research has been conducted to evaluate the potential of radiomics in assessing Ki-67 expression levels in L-NENs. Meyer et al. [16] assessed chest computed tomography (CT) scans in patients’ plain and arterial phases from the same institution and identified several CT texture features correlated with the PI of the lung neuroendocrine Ki-67 antigen to differentiate TC and AC potentially. Another single-institution study analyzing baseline and venous-phase chest CT images obtained using different CT scanner models demonstrated a correlation between texture features and the histological heterogeneity of L-NENs, Ki-67 values, and metastatic foci. Thus, textural features may help assess tumor type and aggressiveness in L-NENs [17]. However, these single-center studies did not establish effective predictive models.

To address these limitations, this study intended to develop a radiographic score using plain and dual-phase enhanced CT fusion images from various hospitals and to create a nomogram for the prediction of Ki-67 PI levels in L-NENs by integrating clinical data and imaging characteristics. Additionally, the study evaluated this model’s performance and generalizability, with the aim of improving clinical decision-making while reducing patient discomfort.

Research methodology

Patients

Ethics approval was obtained from all four participating institutions, with a waiver of informed consent due to the retrospective nature of the study. This study was carried out in full compliance with the ethical principles set forth in the Declaration of Helsinki. The study examined data from consecutive patients with pathologically diagnosed L-NENs at the First Affiliated Hospital of Guangxi Medical University (Center 1) and Guangxi Medical University Cancer Hospital (Center 2) between January 2014 and April 2024, as well as at the Liuzhou Workers’ Hospital (Center 3) and the Guangxi Zhuang Autonomous Region Chest Hospital (Center 4) between January 2019 and April 2024.

The inclusion criteria were: (1) diagnosis of L-NENs through surgical pathology specimens and immunohistochemical examination, and (2) preoperative chest CT scan with dual-phase enhancement. The following were the exclusion criteria: (1) no Ki-67 immunohistochemistry data; (2) poor image quality with severe artifacts; (3) tumor too small (largest diameter < 5 mm); (4) incomplete clinical data; (5) received radiotherapy or chemotherapy before surgery; (6) the tumor with non-small cell carcinoma components.

Patients from Centers 1 to 4 were selected, with 112, 54, 19, and 14 patients from each center respectively. For model development, the 166 patients from Centers 1 and 2 were subsequently divided into training (n = 116) and test (n = 50) cohorts randomly in a 7:3 ratio. The external validation cohort comprised the 33 patients from Centers 3 and 4. Figure 1 depicts the patient selection procedure.

Fig. 1.

Flowchart illustrating the participant selection process

CT examinations

The detailed CT protocol is provided in the Supplementary Material (Table S1, CT Protocol).

Evaluation of radiological features and clinical data

Clinical data, including sex, age, and smoking status, were extracted from each patient’s electronic medical records system. Two chest radiologists with 10 and 20 years of experience in interpreting chest imaging evaluated CT images, which included both contrast-enhanced and non-contrast images. They were blinded to the pathological and clinical outcomes. Consensus was used to settle disagreements. The radiological characteristics listed below were assessed: (1) longest tumor diameter (cm) on axial images; (2) spiculation sign; (3) liquefaction necrosis; (4) calcification; and (5) obstructive pneumonia. A sharp, linear protrusion between the tumor and the surrounding lung parenchyma was identified as the spiculation sign [18].

Immunohistochemical analysis of Ki-67 index

After being fixed in 10% neutral buffered formalin, each sample was routinely dehydrated, embedded in paraffin, and cut into 4-micron-thick slices. Using the Ventana Benchmark Ultra automated IHC staining equipment (Roche Ventana, Inc.), Ki-67 was detected for proliferation by IHC. The percentage of positive tumor cell nuclei was defined as the PI, and cells with brownish nuclei were considered Ki-67-positive. Based on the Ki-67 PI, L-NENs were divided into low (PI ≤ 30%) and high (PI > 30%) expression groups [3].

Tumor segmentation and feature extraction

The Picture Archiving and Communication System (PACS) provided the CT images, which were then converted to Digital Imaging and Communications in Medicine (DICOM) format. 3D Slicer 5.2.2 (https://www.slicer.org) was used to segment the images with a window width of 1500 Hu and window position of -650 Hu. All CT image slices were reviewed for each patient to select the slice with the largest tumor area for measuring the maximum tumor diameter. In the present study, all extraction and selection of CT features were completed using UltraScholar 2.0 (Shukun Network Technology Co., Ltd., Beijing).

To define the volumes of interest (VOIs), 3D slicer was utilized for manual delineation of tumor areas on axial CT images in the plain, arterial, and venous phases of the CT scan. To assess intraobserver reliability, radiologist A resegmented 30 randomly selected cases after two weeks. Furthermore, to assess interobserver reliability, a second radiologist (B) independently segmented these 30 instances without informing either radiologist of the histological findings.

Before feature extraction, the obtained images underwent preprocessing, which involved normalization, modification of image resolution to 1 × 1 × 1 mm³ by B-spline interpolation, and discretization of gray levels using a fixed bin width of 25 Hu. Features were extracted from each phase (non-enhanced, arterial, venous) separately and then concatenated into a combined feature set for subsequent analysis. This approach yielded 1,874 radiomics features in total, comprising 1,500 s-order (texture) characteristics, 360 first-order (histogram) features, and 14 shape features. The workflow is illustrated in Fig. 2. Features with an intraclass correlation coefficient (ICC) and intrareader correlation coefficient, both exceeding 0.75, were deemed to show satisfactory agreement.

Fig. 2.

Details of the radiomic analyses

Feature selection and model building

In the preprocessing phase, features that were constant (zero variance) or had identical values across all samples were removed. Subsequently, redundant features were eliminated using Pearson correlation analysis with a threshold set at |r| > 0.8, whereby features showing high correlation with others were excluded. Following this, LASSO regression was applied to further compress the feature space, with the maximum number of iterations set to 3000, to select the most discriminative features. Finally, a model-based feature selection method was introduced, limiting the maximum number of features to optimize the final feature subset. For model development, the dataset was randomly partitioned into a training set (70%) and a test set (30%) using a fixed random seed of 427 to ensure the reproducibility of the data split. Subsequently, 10-fold cross-validation was employed on the training set to mitigate overfitting risk and enhance the model’s generalizability.

To ensure methodological rigor, the same feature set identified through LASSO was applied to all classifiers for consistent comparison. After that, radiomics signatures were developed using five machine learning algorithms: logistic regression (LR), support vector machine (SVM), extreme gradient boosting (XGBoost), Gaussian Naive Bayes (GaussianNB), and linear support vector machine (LinearSVC). The training cohort underwent ten-fold cross-validation, and the outcomes of the test and external validation cohorts were assessed. The best-performing classifier was selected to construct the radiomics signature (Rad-score).

The clinical-radiomics model was developed by integrating the radiomics features, selected using the consistent parameters (random seed, 70/30 split, 10-fold cross-validation, Pearson threshold of |r| > 0.8, and LASSO settings) applied during the multi-classifier training, with the independent clinical-radiological risk factors identified through multivariate analysis. This unified feature set was then used to train a logistic regression (LR) classifier, and the resulting model was presented as a nomogram for clinical use. The discriminative ability of the clinical-radiomics model was assessed using calibration curves in the training, internal validation, and external validation sets. Decision curve analysis was employed to quantify the net benefit at different threshold probabilities and evaluate the clinical usefulness of the nomogram.

Statistical analysis

Statistical analyses were performed using R (version 4.4.0) and SPSS (version 26.0). Continuous and categorical variables were compared using the Kruskal-Wallis/Mann-Whitney U tests and the χ² test/Fisher’s exact tests, respectively. Independent predictors were identified via univariate and multivariate logistic regression. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), with comparisons made using the DeLong test. Calibration and clinical utility were evaluated using calibration curves and decision curve analysis (DCA), respectively. A P value < 0.05 was considered significant.

Results

Patient characteristics

This retrospective study enrolled 199 patients in total, including 143 males (71.86%) and 56 females (28.14%), with a median age of 59 years (IQR: 51.75–66.25). The high (Ki-67 > 30%) and low (Ki-67 ≤ 30%) Ki-67 status groups included 119 (59.8%) and 80 (40.2%) patients, respectively. Across the training, testing, and external validation cohorts, significant differences were observed only for the largest tumor diameter (P = 0.047) and liquefaction necrosis (P < 0.001), with no significant differences in other characteristics including age, sex, smoking status, calcification, spiculation sign, or obstructive pneumonia, as detailed in Table 1.

Table 1.

Comparison of clinical and radiological characteristics across the training, testing, and external validation cohorts

Characteristics	Training cohort (n = 116)	Testing cohort (n = 50)	External validation cohort (n = 33)	P value
Age, median (Q₁, Q₃)	59.00 (51.75, 66.25)	58.50 (52.00, 66.00)	60.5 (56.5, 72.00)	0.275
Largest tumor diameter, median (Q₁, Q₃)	3.20 (2.20, 5.15)	3.95 (2.10, 5.88)	5.1 (2.85, 6.15)	0.047*
Sex, n (%)
Female	35 (30.17)	11 (22.00)	9 (27.27)	0.523
Male	81 (69.83)	39 (78.00)	24 (72.73)
Smoking status, n (%)				0.878
No	40 (34.48)	16 (32.00)	10 (30.30)
Yes	76 (65.52)	34 (68.00)	23 (69.70)
Liquefaction necrosis, n (%)				< 0.01*
No	113 (97.41)	49 (98.00)	25 (75.76)
Yes	3 (2.59)	1 (2.00)	8 (24.24)
Calcification, n (%)				0.258
No	98 (84.48)	38 (76.00)	24 (72.73)
Yes	18 (15.52)	12 (24.00)	9 (27.27)
Spiculation sign, n (%)				0.799
No	104 (89.66)	46 (92.00)	29 (87.88)
Yes	12 (10.34)	4 (8.00)	4 (12.12)
Obstructive pneumonia, n (%)				0.980
No	67 (57.76)	28 (56.00)	19 (57.58)
Yes	49 (42.24)	22 (44.00)	14 (42.42)

Data are reported as median (Q₁, Q₃) for continuous variables (compared by Kruskal-Wallis test) or n (%) for categorical variables (compared by χ² or Fisher's exact test). * P < 0.05

Age, sex, smoking status, and largest tumor diameter significantly correlated with high Ki-67 expression in univariate analysis (P < 0.05). In the combined multivariate model, the following independent characteristics influenced Ki-67 expression: smoking status (OR: 12.80, P<0.001), largest tumor diameter (OR: 1.79, P<0.001), and age (OR: 1.09, P=0.006) (Table 2).

Table 2.

Univariable and multivariate logistic regression analyses of risk factors for predicting Ki-67 expression in the training cohort

Characteristic	Univariable logistic regression		Multivariable logistic regression
	OR (95% CI)	P-value	OR (95% CI)	P-value
Age	1.09 (1.05 ~ 1.13)	< 0.01*	1.09 (1.04 ~ 1.15)	< 0.01*
Largest tumor diameter (cm)	1.72 (1.39 ~ 2.12)	< 0.01*	1.79 (1.35 ~ 2.37)	< 0.01*
Sex	6.69(3.14 ~ 14.26)	< 0.01	0.80 (0.19 ~ 3.32)	0.757
Smoking history	14.67 (6.65 ~ 32.35)	< 0.01	12.80 (3.39 ~ 48.34)	< 0.01
Liquefaction necrosis	10760310.54 (0.00 ~ Inf)	0.983
Spiculation sign	1.11 (0.38 ~ 3.22)	0.846
Calcification	1.68 (0.72 ~ 3.95)	0.231
Obstructive pneumonia	0.81 (0.45–1.46)	0.487

OR, odds ratio; CI, confidence interval

Note: P values for the difference in clinical characteristics between patients with low and high Ki-67 expression in univariate analysis

P-* Mann–Whitney U test; all others used the χ² test test

Radiomic-based model performance

The radiomics model was constructed using four features with non-zero coefficients selected through the most effective feature selection process (Fig. 3). All four selected features belonged to the gray-level size-zone matrix (GLSZM) family, which quantifies intratumoral heterogeneity by measuring the size of homogeneous connected pixel zones. The logistic regression (LR) algorithm achieved AUCs of 0.912 (95% CI: 0.858–0.965) in the training set, 0.943 (95% CI: 0.887–0.999) in the testing set, and 0.846 (95% CI: 0.786–0.906) in the external validation set, outperforming other traditional machine learning methods (e.g., SVM, XGBoost) across multiple metrics, including AUC, F1-score, and accuracy (Supplementary Material (Table S2)). The model output was converted into a Radscore to represent the relative risk of high Ki-67 expression, and this approach was defined as the radiomics model.

Fig. 3.

The features selected for constructing the radiomics score

Construction and validation of the radiomic nomogram

The combined clinical-radiomics model achieved AUCs of 0.958 (95% CI: 0.925–0.990), 0.930 (0.865–0.995), and 0.911 (0.867–0.955) in training, testing and external validation sets, with good calibration and superior net benefit on decision curve analysis. Figure 4 shows the ROC curves for the clinical, radiomics, and clinical-radiomics nomogram models for the training, testing, and external validation populations. The AUC values, sensitivity, specificity, accuracy, F1 score, positive predictive value (PPV), and negative predictive value (NPV) for the individual models are detailed in Table 3. In all cohorts, the clinical-radiomics nomogram showed the best diagnostic performance for Ki-67 PI expression. The DeLong test indicated that the training cohort’s nomogram model and clinical model had significantly different AUCs. However, there was an insignificant difference between the models in the internal and external validation sets (P > 0.05) and between the AUCs of the radiomics model and the nomogram (P = 0.23).

Fig. 4.

The ROC curves for the nomogram, radiomics signature, and clinical models in the prediction of Ki-67 PI levels in L-NENs throughout training (a), testing (b), and validation cohorts (c). ROC, receiver operating characteristic. PI, proliferative index

Table 3.

Diagnostic performance of the clinical, radiomics, and nomogram models in the training, testing, and external validation cohorts

Models	AUC	95% CI	SEN	SPE	ACC	F1	PPV	NPV
Training cohort (n = 116)
Clinical model	0.898	0.841–0.955	0.812	0.844	0.842	0.848	0.889	0.745
Radiomics model	0.912	0.858–0.965	0.913	0.778	0.860	0.887	0.863	0.853
Nomogram model	0.958	0.925–0.990	0.826	0.955	0.877	0.981	0.966	0.782
Internal validation cohort (n = 50)
Clinical model	0.863	0.749–0.978	0.733	0.800	0.760	0.785	0.846	0.667
Radiomics model	0.943	0.887–0.999	0.900	0.800	0.860	0.885	0.871	0.842
Nomogram model	0.930	0.865–0.995	0.700	0.900	0.780	0.792	0.913	0.667
External validation cohort (n = 33)
Clinical model	0.882	0.828–0.936	0.786	0.831	0.803	0.828	0.875	0.720
Radiomics model	0.846	0.786–0.906	0.765	0.754	0.761	0.794	0.824	0.680
Nomogram model	0.911	0.867–0.955	0.673	0.923	0.773	0.781	0.929	0.652

Note: SEN: Sensitivity; SPE: Specificity; ACC: Accuracy; F1: F1-score; PPV: Positive predictive value; NPV: Negative predictive value

According to the calibration curve analysis, the nomogram model exhibited good agreement with the actual trends in the training, testing, and external validation sets (Fig. 5). The DCA (Fig. 6) also indicated that, compared to the radiomics model and clinical model, the nomogram model tended to show a marginal net benefit in differentiating between individuals with high and low Ki-67 PI levels within the most appropriate threshold probability range. The decision curve analysis demonstrated that the nomogram provided a superior net benefit across a wide range of threshold probabilities (approximately 10%–85%) compared with the clinical or radiomics models alone. For instance, at a threshold probability of 50%, the net benefit of the nomogram was approximately 0.45, indicating its clinical practicality. This threshold range corresponds to plausible clinical decision points. For example, a high predicted probability of elevated Ki-67 might prompt consideration of more extensive preoperative staging, while a low probability could support a less aggressive initial approach.

Fig. 5.

(a) Nomogram for predicting Ki-67 > 30% risk. (b-d) Calibration curves for the training, testing and external validation sets. The close fit of the nomogram line (blue) to the ideal (dashed) indicated high agreement between predicted probabilities and actual outcomes. Radscore, radiomics score

Fig. 6.

Decision curve analysis (DCA) for clinical model, radiomics model, and nomogram model in the training cohort (a), internal validation cohortand (b) and external validation cohort (c). The nomogram (dark blue line) provided the highest net benefit across a wide threshold probability range (10–85%), demonstrating superior clinical utility

Discussion

In this multicenter retrospective study, we developed and rigorously validated a CT-based radiomics nomogram that integrated quantitative imaging features with clinical variables for the preoperative prediction of Ki-67 status in L-NENs. The foremost novelty of our work lies in its successful external validation across independent institutions, a critical step often missing in previous radiomics studies of L-NENs. Prior investigations, such as the single-center studies by Meyer et al. [16] and Cozzi et al. [17], primarily identified correlations between CT texture features and Ki-67 levels but fell short of building and validating a robust predictive model. Our study advances the field by not only constructing a predictive model but also demonstrating its generalizability through external validation, thereby providing stronger evidence for its potential clinical translation. Although our model achieved high AUCs in the training (0.958), internal testing (0.930), and external validation (0.911) cohorts, several methodological considerations merit discussion.

First, radiomics analyses are intrinsically high-dimensional and vulnerable to overfitting when the number of extracted features exceeds the number of clinical events [19]. We attempted to mitigate this risk by removing highly correlated variables (Pearson |r|> 0.8), applying LASSO logistic regression with 10-fold cross-validation, and evaluating the final model in an independent external dataset. Nevertheless, the external validation cohort comprised only 33 patients, producing a 95% confidence interval that remains wide (95% CI: 0.867–0.955) and whose upper bound approached unity, suggesting that the observed performance may be partly influenced by random variation due to the small sample size, consistent with reports of unstable AUC estimates in small external cohorts [20]. Larger, prospective, geographically distinct cohorts scanned with heterogeneous CT platforms are required to confirm generalizability.

Additionally, several methodological aspects should be noted. (1) While applying the same LASSO-selected feature set across different classifiers facilitated a direct performance comparison, it is recognized that different algorithms may have distinct optimal feature subspaces; thus, the relative performance of classifiers should be interpreted within this methodological context. (2) Although 10-fold cross-validation was employed, the feature selection process itself was not nested within the cross-validation folds, which could introduce a degree of optimism in the internal performance estimates on the training set. Nested cross-validation is a recommended refinement for future studies to further reduce this risk. (3) While features from all three CT phases (non-contrast, arterial, venous) were concatenated to maximize information utilization, the study did not quantitatively evaluate the individual contribution or incremental value of each phase to the model’s predictive power.

Second, all four radiomics features retained in our final model belong to the gray-level size-zone matrix (GLSZM) family, which quantifies intratumoral heterogeneity by measuring the size of homogeneous connected pixel zones. Features such as GLSZM_Large Area High Gray Level Emphasis and GLSZM_Zone Variance reflect the presence, distribution, and spatial complexity of high-density cellular regions within the tumor. A high Ki-67 proliferation index is a hallmark of aggressive tumors, often associated with increased cellularity, necrotic regions, and structural disorganization—precisely the textural patterns captured by GLSZM features. Therefore, the dominance of GLSZM features in our model provides a compelling biological rationale: they serve as non-invasive imaging proxies for the underlying tumor proliferation activity and microarchitectural chaos driven by high Ki-67 expression [16, 21]. This aligns with studies in other cancers, reinforcing the link between radiomics heterogeneity and tumor biology. Likewise, recent radiomics models for Ki-67 in sinonasal malignancies [22] and hepatocellular carcinoma [23] confirm the added value of high-order texture features, reinforcing the biological link between intra-tumoral heterogeneity and proliferation. However, radiomic surrogates remain indirect; integration with digital pathology or multiomic profiling will be necessary to uncover mechanistic links between imaging phenotypes and tumor proliferation [11].

Third, our multicenter study design inherently introduces the consideration of feature harmonization to mitigate batch effects. Although advanced algorithms like ComBat were not employed, the rigorous feature selection pipeline we implemented proved effective in extracting robust and generalizable features, as evidenced by the model’s outstanding performance in the external validation set (AUC of 0.911). However, we acknowledge that the incorporation of well-established harmonization techniques such as ComBat in future large-scale multicenter studies is expected to further enhance the model’s reproducibility and robustness across diverse imaging platforms [24].

Fourth, multivariable analysis generated an odds ratio of 12.8 (95% CI: 3.4–48.3) for smoking history, reflecting the marked imbalance in our cohort (85% of smokers versus 28% of non-smokers had Ki-67 > 30%) and the limited number of events, both of which can inflate point estimates. Comparable extreme odds ratios have been reported in small-sample radiomics studies for multiple myeloma [20] and bladder cancer [25], where sparse-event logistic regression was likewise employed. Bootstrapping or Bayesian shrinkage methods should be considered in future work to obtain more stable estimates of effect sizes.

Finally, decision curve analysis indicated a net benefit for the nomogram within the threshold probability range of 10–85% compared to the clinical or radiomics models alone. This threshold range corresponds to plausible clinical decision points. For example, a high predicted probability of elevated Ki-67 might prompt consideration of more extensive preoperative staging, while a low probability could support a less aggressive initial approach. However, DCA does not incorporate real-world costs such as unnecessary biopsies, false-positive downstream imaging, or patient anxiety [26]. Prospective interventional trials are therefore needed to determine whether application of the nomogram truly alters clinical management and to evaluate cost-effectiveness before widespread adoption.

Limitations

Our investigation has several limitations. (1) The retrospective design predisposes to selection and ascertainment bias. (2) The external validation sample was small (n = 33), limiting precision of performance estimates. (3) All L-NEN subtypes were pooled; subtype-specific models might yield better accuracy. (4) Ki-67 assessment followed local pathology protocols without central review. (5) Cost-utility and impact on patient outcomes were not evaluated. (6) Due to the constraints of the analytical software used, confidence intervals for sensitivity, specificity, positive predictive value, and negative predictive value (Table 3) could not be calculated and are thus not reported, limiting the precision assessment of these diagnostic metrics.Larger prospective studies that address these issues are warranted.

Conclusions

In conclusion, the CT-based clinical-radiomics nomogram provides a valuable and non-invasive tool for preoperatively estimating the probability of high Ki-67 expression in patients with lung neuroendocrine neoplasms. This study, through rigorous multicenter development and external validation, demonstrates the potential of integrating quantitative imaging features with clinical variables to inform personalized management strategies. While the current model shows promising accuracy and clinical utility, its integration into routine practice requires further validation. Future work should focus on prospective, large-scale studies in diverse populations to confirm generalizability, and on evaluating the model’s impact on clinical decision-making and patient outcomes. Following successful prospective validation, the development of an accessible online calculation tool will be a key step towards facilitating clinical translation and adoption.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AC

Atypical carcinoids

DICOM

Digital imaging and communications in medicine

DCA

Decision curve analysis

AUC

The area under the curve

CT

Computed tomography

GLSZM

Gray level size zone matrix

IHC

Immunohistochemical

L-NENs

Lung neuroendocrine neoplasms

NET

Neuroendocrine tumors

NEN

Neuroendocrine neoplasms

NEC

Neuroendocrine carcinomas

ROC

Receiver operating curve

PACS

Picture archiving and communication system

PI

Proliferation index

SCLC

Small-cell lung cancer

LCNEC

Large-cell neuroendocrine carcinomas of the lung

TC

Typical carcinoids

VOI

Volume of interest

Author contributions

Xiao Pan, Yanni Zou, Tao Li, Peng Peng, and Wenhua Zhao contributed to the study conception and design. Material preparation and data collection were performed by Xiao Pan, Yanni Zou, Xiaoxiao Huang, and Quan Zhang. Data analyses were performed by Xiao Pan and Jing Hu. The first draft of the manuscript was written by Xiao Pan, Yanni Zou. Jing Hu, a senior clinical scientist at Shukun (Beijing) Technology Co., Ltd., contributed to statistical analysis and manuscript editing. The company had no control over the study. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by ① the Joint Regional Epidemic Disease Research Grant (project No. 2022JJA140455), titled ‘A quantitative MRI study of pituitary iron deposition in patients with Mediterranean anemia’; ② the Provincial Science and Technology Program (project No. S2022072), titled ‘A quantitative study on the correlation between hepatic iron deposition and glucose metabolism in Mediterranean anemia using multimodal MRI’.

Data availability

The data supporting this study are available from the corresponding author upon reasonable request but may be restricted to protect patient confidentiality and comply with ethical guidelines, given the study’s retrospective design.

Declarations

Ethics approval and consent to participate

This retrospective study, conducted in accordance with the Declaration of Helsinki, received approval from the institutional review boards of The First Affiliated Hospital of Guangxi Medical University (approval No. 2025-E0595), Guangxi Medical University Cancer Hospital (approval No. KY-2022-301), Liuzhou Workers’ Hospital (approval No. KY2025614), and The Chest Hospital of Guangxi Zhuang Autonomous Region (approval No. 2023-S004-01), each of which waived the requirement for informed patient consent.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.