Abstract
Background
Lung cancer is a leading cause of cancer-related death worldwide, and accurate differentiation of pulmonary nodules is crucial for clinical management. Low-dose and multiphasic computed tomography (CT) provides improved nodule detection, but the incremental diagnostic value of contrast-enhanced phases remains unclear. The objective of this study was to establish a prediction model based on clinical and CT data from three phases (noncontrast phase, arterial phase, and venous phase) for differentiating benign and malignant pulmonary nodules.
Methods
A total of 863 patients from multiple centers with benign or malignant pulmonary nodules pathologically confirmed between 2015 and 2024 were retrospectively analyzed. Data from The First Affiliated Hospital of Bengbu Medical University and Anqing Municipal Hospital (Anhui, China), constituted the training cohort (n=470), data from Tongde Hospital of Zhejiang Province Affiliated to Zhejiang Chinese Medical University and Hangzhou TCM Hospital Affiliated to Zhejiang Chinese Medical University (Zhejiang, China), were used as the test cohort (n=205), and data from an additional center, Hangzhou TCM Hospital Affiliated to Zhejiang Chinese Medical University (Hangzhou, Zhejiang, China), served as the validation cohort (n=188). According to different CT scanning phases, the data were divided into the following four groups, from which models were constructed: noncontrast phase, arterial phase, venous phase, and noncontrast phase each combined with the enhanced dual phase. Candidate variables were screened via univariate analysis and subsequently entered into multivariable logistic regression to identify independent predictors for model construction. Model performance was evaluated via receiver operating characteristic analysis and the area under the curve (AUC).
Results
A total of 863 patients with pulmonary nodules were included, comprising 518 malignant nodules and 345 benign nodules. The final models incorporated clinical and imaging predictors, including age, surgical history, hepatitis or liver cirrhosis, interstitial lung disease, bronchiectasis, obstructive pneumonia, mixed ground-glass opacity, lobulation, pleural retraction, minimum CT attenuation value, and mean CT attenuation value. Model 1 retained 11 predictors, whereas models 2–4 retained 10 predictors with identical predictor sets. In the validation cohort, the AUC, sensitivity, and specificity values for model 1 were 0.909, 0.852, and 0.740, respectively; those for models 2 and 4 were 0.904, 0.755 and 0.849, respectively; and those for model 3 were 0.904, 0.870, and 0.712, respectively. There was no significant difference in the diagnostic performance of the four models (P>0.05).
Conclusions
The four prediction models established based on clinical and three-phase CT data all demonstrated high diagnostic efficiency in differentiating benign and malignant pulmonary nodules.
Keywords: Computed tomography (CT), pulmonary nodules, benign and malignant, different phase identification model
Introduction
Lung cancer remains the most commonly diagnosed malignancy and one of the leading causes of cancer-related death worldwide, representing a major public health burden, particularly in China (1). In 2022, lung cancer ranked first in both incidence and mortality among all cancers in China, with an incidence rate of 75.13 per 100,000 and a mortality rate of 51.94 per 100,000, making it the leading cause of cancer-related death (2). Therefore, improving early detection and diagnostic accuracy is critical for reducing lung cancer-related mortality.
In recent years, the widespread implementation of low-dose computed tomography (CT) for lung cancer screening, along with advances in artificial intelligence-assisted diagnostic tools, has led to a substantial increase in the detection of pulmonary nodules (3,4). However, the majority of these nodules are benign (5,6). Although increased detection facilitates early diagnosis, it also introduces new clinical challenges. False-positive findings may result in unnecessary invasive procedures, overdiagnosis, and overtreatment (7,8), while benign nodules may be misclassified as malignant, leading to unwarranted surgical interventions (9). Consequently, accurate differentiation between benign and malignant pulmonary nodules has become a critical issue in clinical practice.
Chest spiral CT is currently the primary imaging modality for the evaluation of pulmonary nodules. Noncontrast CT combined with three-dimensional reconstruction can comprehensively depict morphological characteristics, including size, shape, margins, and relationships with adjacent structures (10,11). On this basis, contrast-enhanced CT provides additional functional information by reflecting tumor vascularity, thus further informing the differentiation of benign from malignant nodules (12,13). However, the incremental diagnostic value provided by the different CT phases has not been sufficiently determined.
In recent years, numerous studies have developed predictive models integrating clinical and CT imaging features to improve diagnostic accuracy (14-16). Although these approaches have achieved encouraging results, several limitations remain. First, some studies (14,15) relied on single-phase CT features without fully leveraging multiphasic imaging information. Second, the comparative diagnostic performance of different CT phases (noncontrast, arterial, and venous) and their combinations have not been systematically evaluated, and no consensus has been reached. In addition, there are few descriptive stratified analyses based on arterial-phase CT enhancement values reported in the literature. As a result, there remains a lack of comprehensive evidence based on integrated clinical and multiphasic CT features for model construction and performance comparison.
Therefore, we conducted a multicenter cohort study with the aim of generating findings with greater generalizability and robustness. The primary objective was to systematically evaluate the diagnostic performance of different CT phases (noncontrast, arterial, and venous) and their combinations in differentiating benign from malignant pulmonary nodules. Based on this framework, clinical and imaging features from each CT phase were integrated to construct phase-specific binary logistic regression models, as well as a combined multiphasic model incorporating information from all phases, enabling comparison of different diagnostic strategies. The study population was stratified according to arterial-phase CT enhancement levels, and the proportions of benign and malignant lesions within each stratum were descriptively analyzed to preliminarily examine the distribution patterns between enhancement degree and lesion nature. The novelty of this study lies in that, beyond systematically comparing the diagnostic performance of models based on different CT phases and their combinations, we further incorporated enhancement-based stratified descriptions, thereby providing reference data for future large-scale and more refined quantitative modeling. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0120/rc).
Methods
Patients
This multicenter retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by The First Affiliated Hospital of Bengbu Medical University (approval No. LWSL202300440), Anqing Municipal Hospital (approval No. 83230471), Tongde Hospital of Zhejiang Province (approval No. MR-33-24-041194), Hangzhou Hospital of Traditional Chinese Medicine (approval No. 2025KLL007), and Taizhou Municipal Hospital (approval No. LWYJ2023059). The requirement for informed consent was waived due to the retrospective nature of the analysis.
Patients with pulmonary nodules who underwent routine preoperative chest CT and dual-phase enhanced CT scans from 2015 to 2024 were retrospectively enrolled. The preoperative CT images and clinical data were collected and analyzed. A total of 1,569 patients met the inclusion criteria and were further screened according to the following exclusion criteria: incomplete clinical data, preoperative treatment, poor CT scan image quality, and a preoperative CT scan time of more than 1 month. Data from The First Affiliated Hospital of Bengbu Medical University and Anqing Municipal Hospital (Anhui, China) were used as the training cohort (n=470), data from Tongde Hospital of Zhejiang Province Affiliated to Zhejiang Chinese Medical University and Hangzhou TCM Hospital Affiliated to Zhejiang Chinese Medical University (Zhejiang, China) constituted the test cohort (n=205), and data from Taizhou Municipal Hospital (Zhejiang, China) were used the external validation cohort (n=188). A total of 863 patients were included for the subsequent analyses (Figure 1).
Figure 1.
Flowchart of patient inclusion. Center 1+2: The First Affiliated Hospital of Bengbu Medical University and Anqing Municipal Hospital (Anhui, China); Center 3+4: Tongde Hospital of Zhejiang Province Affiliated to Zhejiang Chinese Medical University and Hangzhou TCM Hospital Affiliated to Zhejiang Chinese Medical University (Zhejiang, China); Center 5: Taizhou Municipal Hospital (Zhejiang, China). CT, computed tomography.
CT image acquisition
Chest CT imaging was performed with a variety of scanners, including the Optima CT680, Optima CT540, BrightSpeed 16 (GE HealthCare, Chicago, IL, USA), SOMATOM Definition Flash CT, SOMATOM Force 192, SOMATOM Sensation 16 (Siemens Healthineers, Erlangen, Germany), Brilliance iCT SP 128, and Incisive CT (Philips, Amsterdam, the Netherlands). The tube voltage was set to 120 kV, and the tube current was automatically adjusted from 150 to 300 mA. Both layer thickness and layer spacing were 5 mm. Thin-layer reconstructions were performed through use of standard reconstruction algorithms, with slice thickness and slice spacing ranging from 0.5 to 1.25 mm. All patients underwent conventional and enhanced dual-phase scanning from the apex to the base of the lung. Contrast agent was injected via the antecubital vein with a high-pressure syringe at a dose of 1.5 mL/kg and a flow rate of 2.5–3.0 mL/s. Dual-phase scanning was performed in a 30-s arterial phase and a 60-s venous phase after injection of contrast agent. All CT images were displayed with standard lung and mediastinal window settings, with the lung window width and level being 1,200 to 1,500 Hounsfield unit (HU) and –600 to –400 HU, respectively, and the mediastinal window width and level being 400 and 40 HU, respectively. Although formal statistical correction for interscanner differences was not performed in this study, standardized region of interest (ROI) delineation and CT value measurement were conducted at the identical slice and region of each lesion, and these settings were replicated across different CT phases to minimize the impact of heterogeneity.
Analysis of images
Image evaluation and interobserver agreement
Two radiologists with 5 and 15 years of diagnostic experience independently analyzed and evaluated chest CT images according to the same standard. If there was disagreement, a third senior radiologist participated in the discussion until a consensus was reached. The radiologists were blinded to the pathological results during image evaluation. To ensure the consistency of different types of data, the intraclass correlation coefficient (ICC) and Cohen’ kappa were used to evaluate the continuous and categorical variables, respectively. An ICC >0.75 indicated good consistency, and a kappa value (κ) >0.8 indicated high interrater reliability.
ROI delineation and quantitative CT measurement
In noncontrast CT images, ROIs were manually delineated on the slice showing the maximum cross-sectional area of the lesion. The ROI boundaries were carefully defined according to the natural anatomic interface between the lesion and surrounding normal tissues. During segmentation, areas of necrosis, calcification, vessels, bronchi, and obvious normal lung parenchyma were avoided as much as possible to minimize the influence of nontumor components on quantitative measurements. In contrast-enhanced scans (arterial and venous phases), the same ROIs were copied and applied to the corresponding slice as in the noncontrast phase for analysis, ensuring consistency across different phases. In cases of slight positional shifts or respiratory motion differences, visual alignment based on anatomical landmarks was performed prior to ROI replication to improve measurement accuracy and consistency. All CT quantitative measurements were performed under standardized protocols to ensure comparability across different cases and imaging phases.
Clinical variables
Clinical data examined in this study included sex, age, smoking status, surgical history, high blood pressure (HBP), diabetes mellitus (DM), cardiovascular and cerebrovascular diseases, hepatitis or liver cirrhosis, and tumor markers.
Imaging features
The chest CT features examined included the following: (I) morphological features, including morphology, location, mixed ground-glass opacity (mGGO), lobulation, and spiculation; (II) associated findings, including pulmonary emphysema or lung bullae, bronchiectasis, air bronchogram, pleural retraction, lung ventilation and perfusion inhomogeneity, interstitial lung disease, and obstructive pneumonia; (III) structural changes, including cavity, cystic degeneration, and calcification; (IV) quantitative parameters, including long diameter (LD), short diameter (SD), maximum computed tomography attenuation value (CTmax), minimum computed tomography attenuation value (CTmin), mean computed tomography attenuation value (CTmean), standard deviation of computed tomography attenuation value (CTsd), computed tomography attenuation value in the arterial phase (CTa), computed tomography attenuation value in the venous phase (CTv); and (V) enhancement parameters, including computed tomography arterial-phase enhancement value (CTap; CTap = CTa − CTmean) and computed tomography venous-phase enhancement value (CTvp; CTvp = CTv − CTmean).
There were missing data for any of the included variables, including clinical characteristics, CT imaging features, or pathological outcomes. All cases were complete cases, and no additional missing data handling procedures were required. Data for all variables were collected according to predefined criteria for subsequent statistical analysis and model development.
Outcome
The outcome of this study was the pathological classification of pulmonary nodules (benign vs. malignant) as confirmed by postoperative histopathological examination. As the outcome assessment was based on objective pathological findings, no additional blinding procedure was considered necessary.
Feature selection and model construction
The model was developed through use of a data-driven approach based on statistical selection procedures. Feature selection was initially performed in the training set via univariate analysis to identify candidate variables associated with the outcome, with a significance threshold of P<0.05. Variables with statistical significance were subsequently included in a multivariable logistic regression analysis. All selected variables were entered into the model via the enter method to identify independent risk factors. Multicollinearity among variables was assessed according to the variance inflation factor (VIF), and variables with VIF >5 were excluded. Four CT-phase-specific models were constructed under the same feature selection strategy; however, the variables included in each model were phase-specific. In addition to clinical and conventional CT features, the noncontrast model included CTmax, CTmin, CTmean, and CTsd. The arterial-phase model further incorporated CTa and CTap, while the venous-phase model included CTv and CTvp. The combined noncontrast and contrast-enhanced model included all CT parameters (CTmax, CTmin, CTmean, CTsd, CTa, CTap, CTv, and CTvp). The purpose of constructing multiple models was to compare the diagnostic performance across different CT phases and to evaluate whether contrast-enhanced CT provides additional diagnostic value over noncontrast CT (Figure 2).
Figure 2.
Flowchart of model construction. CE-CT, contrast-enhanced computed tomography; CT, computed tomography; CTa, computed tomography attenuation value in the arterial phase; CTap, computed tomography arterial-phase enhancement value (CTap = CTa − CTmean); CTmax, maximum computed tomography attenuation value; CTmean, mean computed tomography attenuation value; CTmin, minimum computed tomography attenuation value; CTsd, standard deviation of computed tomography attenuation value; CTv, computed tomography attenuation value in the venous phase; CTvp, computed tomography venous-phase enhancement value (CTvp = CTv − CTmean); ROC, receiver operating characteristic.
Statistical analysis
SPSS 31.0 statistical software (IBM Corp., Armonk, NY, USA) was used for data analysis. Categorical variables are expressed as frequencies and percentages and were analyzed with the Pearson chi-squared test, Spearman chi-squared test, or Fisher exact test. Continuous variables are expressed as the mean ± standard deviation or as the median and were analyzed with independent samples t-test if in accordance with a normal distribution; otherwise, the Mann-Whitney U test was applied. The training cohort was used for model development, the testing cohort was used for preliminary evaluation of model performance, and the validation cohort was used for independent assessment of the model’s generalizability and robustness. Model performance was evaluated from two aspects: discrimination and calibration. Discrimination was assessed according to receiver operating characteristic (ROC) curves, and the area under the curve (AUC) was calculated to quantify classification performance. The optimal cutoff value was determined based on the maximum Youden index derived from the ROC curve in the training cohort. This cutoff was then fixed and applied to the test and validation cohorts to calculate sensitivity, specificity, and accuracy, as well as to construct confusion matrices. The calibration curve was finally used to assess the calibration performance of the predictive model. Differences in AUC between models were compared via the DeLong test. A two-sided P value <0.05 was considered statistically significant. No hyperparameter tuning was performed in this study. The original model established in the training cohort was directly applied to the test and validation cohorts without any model updates or recalibration.
To further investigate the association between CTap and the risk of malignancy, a stratified analysis was performed. Based on previously reported enhancement ranges (e.g., 20–60 HU) and the distribution characteristics of the data, three different stratification schemes were defined: scheme 1 (<20, 20–40, and >40 HU), scheme 2 (<20, 20–60, and >60 HU), and scheme 3 (<30, 30–50, and >50 HU). For each cohort (training, test, and external validation), the proportion of malignant nodules within each CTap category was calculated. The distribution of malignancy risk across different strata was then compared to evaluate the stability and discriminative ability of each stratification strategy.
Results
Clinical and pathological characteristics
A total of 863 patients with benign and malignant pulmonary nodules were enrolled in this study, including 522 males and 341 females. There were 518 cases with malignant nodules, whose average age was 65.25±10.02 years. There were 345 cases with benign nodules, whose average age was 58.02±12.24 years. A total of 470 patients from The First Affiliated Hospital of Bengbu Medical University and Anqing Municipal Hospital (Anhui, China) were included in the training cohort, including 286 malignant nodules and 184 benign nodules. A total of 205 cases from Tongde Hospital of Zhejiang Province Affiliated to Zhejiang Chinese Medical University and Hangzhou TCM Hospital Affiliated to Zhejiang Chinese Medical University (Zhejiang, China) were used in the test cohort, including 117 malignant nodules and 88 benign nodules. A total of 188 cases from Taizhou Municipal Hospital (Zhejiang, China) were used in the validation cohort, including 115 cases of malignant pulmonary nodules and 73 cases of benign pulmonary nodules. The demographic characteristics of the patients included in this study are detailed in Table 1. The detailed pathological spectrum of pulmonary nodules is provided in Table S1. Among the 345 benign nodules, inflammatory lesions accounted for the majority (237/345, 68.7%), including granulomatous inflammation, fungal infections, tuberculosis, and organizing pneumonia. Benign neoplasms and other noninflammatory nodules accounted for the remaining cases. The most common type of malignant nodule was invasive adenocarcinoma, followed by squamous cell carcinoma and other histological subtypes.
Table 1. Qualitative CT characteristics of the three patient cohorts.
| Characteristic | Benign tumors (n=345) | Malignant tumors (n=518) | P value (training cohort) | |||||
|---|---|---|---|---|---|---|---|---|
| Training cohort (n=184) | Test cohort (n=88) | Validation cohort (n=73) | Training cohort (n=286) | Test cohort (n=117) | Validation cohort (n=115) | |||
| Sex | 0.954 | |||||||
| Female | 69 (37.50) | 51 (58.00) | 26 (35.60) | 108 (37.80) | 44 (37.60) | 43 (37.40) | ||
| Male | 115 (62.50) | 37 (42.00) | 47 (64.40) | 178 (62.20) | 73 (62.40) | 72 (62.60) | ||
| Smoke | 0.037* | |||||||
| Never | 148 (80.40) | 70 (79.50) | 56 (76.70) | 201 (70.30) | 92 (78.60) | 79 (68.70) | ||
| Current | 24 (13.00) | 16 (18.20) | 13 (17.80) | 50 (17.50) | 14 (12.00) | 24 (20.90) | ||
| Former | 12 (6.50) | 2 (2.30) | 4 (5.50) | 35 (12.20) | 11 (9.40) | 12 (10.40) | ||
| Surgical history | 0.002* | |||||||
| No | 122 (66.30) | 60 (68.20) | 58 (79.50) | 230 (80.40) | 107 (91.50) | 97 (84.40) | ||
| Lung cancer | 12 (6.50) | 14 (15.90) | 11 (15.10) | 14 (4.90) | 3 (2.60) | 12 (10.40) | ||
| Other | 50 (27.20) | 14 (15.90) | 4 (5.50) | 42 (14.70) | 7 (6.00) | 6 (5.20) | ||
| HBP | 0.541 | |||||||
| Yes | 48 (26.10) | 23 (26.10) | 22 (30.10) | 82 (28.70) | 43 (36.80) | 39 (33.90) | ||
| DM | 0.004* | |||||||
| Yes | 29 (15.80) | 13 (14.80) | 13 (17.80) | 21 (7.30) | 16 (13.70) | 11 (9.60) | ||
| Cardiovascular and cerebrovascular diseases | 0.002* | |||||||
| Yes | 7 (3.80) | 1 (1.10) | 5 (6.80) | 34 (11.90) | 28 (23.90) | 11 (9.60) | ||
| Hepatitis/liver cirrhosis | 0.012* | |||||||
| Yes | 10 (5.40) | 2 (2.30) | 8 (11.00) | 4 (1.40) | 6 (5.10) | 1 (0.90) | ||
| Pulmonary emphysema or lung bullae | <0.001* | |||||||
| Yes | 43 (23.40) | 15 (17.00) | 11 (15.10) | 115 (40.20) | 43 (36.80) | 37 (32.20) | ||
| Bronchiectasis | 0.002* | |||||||
| Yes | 35 (19.00) | 2 (2.30) | 2 (2.70) | 26 (9.10) | 7 (6.00) | 4 (3.50) | ||
| Ventilation-perfusion mismatch | 0.753 | |||||||
| Yes | 19 (10.30) | 7 (8.00) | 10 (13.70) | 27 (9.40) | 5 (4.30) | 11 (9.60) | ||
| Interstitial lung disease | 0.032* | |||||||
| Yes | 22 (12.00) | 1 (1.10) | 1 (1.40) | 18 (6.30) | 10 (8.50) | 11 (9.60) | ||
| Tumor markers | 0.382 | |||||||
| Yes | 78 (42.40) | 17 (19.30) | 16 (21.90) | 133 (46.50) | 57 (48.70) | 27 (23.50) | ||
| Location | 0.016* | |||||||
| Right upper lobe | 42 (22.80) | 18 (20.50) | 19 (26.00) | 91 (31.80) | 41 (35.00) | 27 (23.50) | ||
| Right middle lobe | 19 (10.30) | 9 (10.20) | 5 (6.80) | 17 (5.90) | 4 (3.40) | 7 (6.10) | ||
| Right lower lobe | 47 (25.50) | 20 (22.70) | 17 (23.30) | 50 (17.50) | 24 (20.50) | 29 (25.20) | ||
| Left upper lobe | 38 (20.70) | 20 (22.70) | 16 (21.90) | 72 (25.20) | 30 (25.60) | 34 (29.60) | ||
| Left lower lobe | 38 (20.70) | 21 (23.90) | 16 (21.90) | 56 (19.60) | 18 (15.40) | 18 (15.70) | ||
| Morphology | <0.001* | |||||||
| Round | 19 (10.30) | 27 (30.70) | 4 (5.50) | 0 (0.00) | 1 (0.90) | 1 (0.90) | ||
| Oval | 18 (9.80) | 22 (25.00) | 13 (17.80) | 4 (1.40) | 3 (2.60) | 5 (4.30) | ||
| Irregular | 147 (79.90) | 39 (44.30) | 56 (76.70) | 282 (98.60) | 113 (96.60) | 109 (94.80) | ||
| Lobulation | <0.001* | |||||||
| Yes | 137 (74.50) | 33 (37.50) | 49 (67.10) | 270 (94.40) | 110 (94.0) | 106 (92.20) | ||
| Spiculation | <0.001* | |||||||
| No | 101 (54.90) | 63 (71.60) | 45 (61.60) | 71 (24.80) | 21 (18.00) | 57 (49.60) | ||
| Short (≤5 mm) | 50 (27.20) | 3 (3.40) | 10 (13.70) | 121 (42.30) | 44 (37.60) | 43 (37.40) | ||
| Long (>5 mm) | 33 (17.90) | 22 (25.00) | 18 (24.70) | 94 (32.90) | 52 (44.40) | 15 (13.00) | ||
| Air space | 0.247 | |||||||
| 0 | 159 (86.40) | 79 (89.80) | 62 (84.90) | 234 (81.80) | 101 (86.30) | 105 (91.30) | ||
| ≤5 | 11 (6.00) | 5 (5.70) | 4 (5.50) | 33 (11.50) | 9 (7.70) | 7 (6.10) | ||
| >5 | 9 (4.90) | 0 (0.00) | 5 (6.80) | 12 (4.20) | 2 (1.70) | 1 (0.90) | ||
| Cavity | 5 (2.70) | 4 (4.50) | 2 (2.70) | 7 (2.40) | 5 (4.30) | 2 (1.70) | ||
| Air bronchogram | <0.001* | |||||||
| No | 145 (78.80) | 81 (92.00) | 47 (64.40) | 146 (51.10) | 27 (23.10) | 69 (60.00) | ||
| Without tracheal deformation | 25 (13.60) | 4 (4.60) | 23 (31.50) | 23 (8.00) | 7 (6.00) | 19 (16.50) | ||
| Accompanied by tracheal deformation or traction | 13 (7.10) | 2 (2.30) | 3 (4.10) | 76 (26.60) | 42 (35.90) | 15 (13.00) | ||
| Bronchiectasis or ectopic displacement due to traction | 1 (0.50) | 1 (1.10) | 0 (0.00) | 41 (14.30) | 41 (35.00) | 12 (10.40) | ||
| Obstructive inflammation | <0.001* | |||||||
| Yes | 7 (3.80) | 3 (3.40) | 0 (0.00) | 99 (34.60) | 69 (59.00) | 23 (20.00) | ||
| Pleural retraction | <0.001* | |||||||
| Type 0 | 80 (43.50) | 36 (40.90) | 11 (15.10) | 44 (15.40) | 30 (25.60) | 27 (23.50) | ||
| Type I | 45 (24.50) | 31 (35.20) | 35 (47.90) | 60 (21.00) | 11 (9.40) | 40 (34.80) | ||
| Type II | 7 (3.80) | 5 (5.70) | 20 (27.40) | 32 (11.20) | 5 (4.30) | 8 (7.00) | ||
| Type III | 22 (12.00) | 4 (4.60) | 0 (0.00) | 91 (31.80) | 52 (44.40) | 7 (6.10) | ||
| Type IV | 30 (16.30) | 12 (13.60) | 7 (9.60) | 59 (20.60) | 19 (16.20) | 33 (28.70) | ||
| Calcification | 0.084 | |||||||
| Yes | 23 (12.50) | 13 (14.80) | 2 (2.70) | 22 (7.70) | 6 (5.10) | 9 (7.80) | ||
| mGGO | 0.016* | |||||||
| ≤25% | 2 (1.10) | 0 (0.00) | 1 (1.40) | 9 (3.10) | 1 (0.90) | 6 (5.20) | ||
| ≤50% | 0 (0.00) | 0 (0.00) | 0 (0.00) | 6 (2.10) | 0 (0.00) | 7 (6.10) | ||
| ≤75% | 4 (2.20) | 1 (1.10) | 0 (0.00) | 16 (5.60) | 4 (3.40) | 6 (5.20) | ||
| <100% | 13 (7.10) | 2 (2.30) | 0 (0.00) | 30 (10.50) | 6 (5.10) | 6 (5.20) | ||
| 100% | 165 (89.70) | 85 (96.60) | 72 (98.60) | 225 (78.70) | 106 (90.60) | 90 (78.30) | ||
| Cystic airspace | <0.001* | |||||||
| Yes | 11 (6.00) | 6 (6.80) | 4 (5.50) | 79 (27.60) | 49 (41.90) | 19 (16.50) | ||
Data are presented as frequencies (percentages) analyzed with the Pearson χ2 test, Spearman χ2 test, or Fisher exact test. Pleural retraction: type 0, no contact traction; type I, with contact between pulmonary nodules and pleura; type II, with traction and no displacement; type III, with traction and displacement; type IV, with contact and traction. *, statistically significant P value (P<0.05). CT, computed tomography; DM, diabetes mellitus; HBP, high blood pressure; mGGO, mixed ground-glass opacity.
Univariate and multivariable analyses
Univariate analysis indicated that age, smoking status, surgical history, DM, cardiovascular and cerebrovascular diseases, hepatitis or liver cirrhosis, pulmonary emphysema or lung bullae, interstitial lung disease, obstructive pneumonia, bronchiectasis, mGGO, location, shape, lobulation, spiculation, air bronchogram, pleural retraction, cystic degeneration, LD, SD, CTmax, CTmin, CTmean, CTsd, CTap, and CTvp were significantly associated with the outcome (P<0.05). After multivariable analysis, the factors that remained as independent predictors were age, surgical history, hepatitis or liver cirrhosis, interstitial lung disease, bronchiectasis, obstructive pneumonia, mGGO, lobulation, pleural retraction, CTmin, and CTmean (Figure 3). The results of the multivariable logistic regression analysis are presented as odds ratios with the 95% confidence interval (CI) in Table S2.
Figure 3.
The independent risk factors of the four models were obtained via binary logistic regression analysis. (A-C) The distribution of age, CTmin, and CTmean in benign and malignant pulmonary nodules, respectively. (D) The proportion of patients in the benign and malignant pulmonary nodule groups with lobulation, surgical history, hepatitis or liver cirrhosis, bronchiectasis, interstitial lung disease, and obstructive pulmonary disease. (E) The proportion of nodule type (based on solid component, including ≤25%, ≤50%, ≤75%, ≤100%, and 100%) in benign and malignant pulmonary nodules. (F) The proportion of different pleural retraction types in benign and malignant pulmonary nodules (including no pleural retraction, type I with contact between pulmonary nodules and pleura, type II with traction and no displacement, type III with traction and displacement, and type IV with contact and traction). CTmean, mean computed tomography attenuation value; CTmin, minimum computed tomography attenuation value; mGGO, mixed ground-glass opacity.
Qualitative CT features
Of the 23 classification features analyzed, six features (sex, HBP, tumor markers, lung ventilation and perfusion heterogeneity, cavitation, and calcification) did not show statistically significant differences between benign and malignant nodules, while the remaining 17 features differed significantly (P<0.05). Among them, shape irregularity, lobulation, spiculation, air bronchogram, obstructive pneumonia, pleural retraction, cystic change were significantly more common observed in malignant nodules than in benign nodules (P<0.001) (Table 1). For qualitative imaging features, Cohen kappa values indicated good-to-excellent agreement (κ=0.81–0.88; Table S3), with calcification showing the highest consistency (κ=0.88).
CT quantitative features
Except for CTa and CTv, all continuous features were significantly different between malignant and benign pulmonary nodules (P<0.05). The mean age of patients with malignant pulmonary nodules was significantly higher than that of patients with benign nodules (64.61±10.32 vs. 56.05±11.76 years) (P<0.001). The median LD and SD of malignant pulmonary nodules [LD: 27.00 mm, interquartile range (IQR), 17.00–40.00 mm; SD: 21.00 mm, IQR, 14.00–31.40 mm] were larger than those of benign pulmonary nodules (LD: 20.00 mm, IQR, 15.00–28.50 mm; SD: 15.00 mm, IQR, 11.00–21.50 mm). However, the enhanced CT values of malignant pulmonary nodules (CTa: 48.90 HU, IQR, 29.60–66.40 HU; CTv: 60.80 HU, IQR, 45.20–78.00 HU) were lower than those of benign pulmonary nodules (CTa: 52.40 HU, IQR, 33.30–69.85 HU; CTv: 64.70 HU, IQR, 38.80–76.45 HU) (Table 2 and Figure 4). The interobserver agreement for quantitative CT measurements was good to excellent, with ICC values ranging from 0.76 to 0.85 for all variables (Table S4). The highest agreement was observed for CTmin (ICC =0.85), followed by CTmean (ICC =0.84).
Table 2. Quantitative CT characteristics of the three patient cohorts.
| Characteristic | Benign tumors (n=345) | Malignant tumors (n=518) | P value (training cohort) | |||||
|---|---|---|---|---|---|---|---|---|
| Training cohort (n=184) | Test cohort (n=88) | Validation cohort (n=73) | Training cohort (n=286) | Test cohort (n=117) | Validation cohort (n=115) | |||
| Age (years) | 58.84±12.41 | 57.93±12.21 | 56.05±11.76 | 66.09±9.37 | 63.64±10.69 | 64.61±10.32 | <0.001* | |
| LD (mm) | 17.00 (12.00, 30.00) | 15.50 (11.00, 24.50) | 20.00 (15.00, 28.50) | 31.00 (22.00, 44.00) | 33.00 (25.00, 52.00) | 27.00 (17.00, 40.00) | <0.001* | |
| SD (mm) | 13.00 (9.00, 21.00) | 11.50 (9.00, 17.00) | 15.00 (11.00, 21.50) | 24.00 (16.00, 32.75) | 25.00 (17.50, 36.00) | 21.00 (14.00, 31.40) | <0.001* | |
| CTmax (HU) | 162.50 (110.25, 245.00) | 113.50 (78.25, 151.50) | 186.00 (126.00, 263.00) | 81.00 (58.25, 106.00) | 77.00 (62.00, 101.00) | 141.00 (94.00, 210.00) | <0.001* | |
| CTmin (HU) | −70.00 (−138.00, −18.25) | −47.00 (−75.25, −15.50) | −101.00 (−165.50, −51.50) | −11.00 (−40.00, 10.00) | 3.00 (−11.00, 14.00) | −101.00 (−200.00, −16.00) | <0.001* | |
| CTmean (HU) | 43.45 (28.18, 57.93) | 33.45 (21.35, 44.25) | 41.20 (30.15, 50.05) | 36.10 (21.68, 47.00) | 38.70 (30.75, 49.50) | 32.00 (16.40, 42.90) | <0.001* | |
| CTsd (HU) | 67.50 (36.60, 98.03) | 40.40 (25.18, 61.38) | 72.10 (53.95, 103.25) | 22.00 (17.00, 34.38) | 22.00 (16.00, 30.00) | 58.00 (35.00, 82.90) | <0.001* | |
| CTa (HU) | 56.90 (28.50, 77.00) | 40.75 (27.95, 59.65) | 52.40 (33.30, 69.85) | 57.85 (40.80, 71.00) | 59.00 (51.25, 75.65) | 48.90 (29.60, 66.40) | 0.930 | |
| CTap (HU) | 13.40 (−5.50, 33.75) | 10.50 (−1.53, 21.38) | 10.00 (0.40, 24.30) | 20.40 (8.50, 36.48) | 20.00 (8.60, 32.65) | 17.00 (6.80, 29.50) | <0.001* | |
| CTv (HU) | 61.15 (34.28, 84.60) | 50.40 (38.00, 69.23) | 64.70 (38.80, 76.45) | 62.90 (47.73, 73.98) | 62.70 (50.55, 74.05) | 60.80 (45.20, 78.00) | 0.953 | |
| CTvp (HU) | 16.25 (−2.78, 43.40) | 14.90 (7.73, 32.45) | 22.50 (5.45, 32.50) | 26.00 (12.08, 40.15) | 22.00 (9.00, 34.90) | 28.60 (13.00, 46.10) | 0.002* | |
Data are presented as mean ± standard deviation or median (interquartile range). If in accordance with normal distribution, the independent samples t-test was used. For data with a nonnormal distribution, the Mann-Whitney U test was used. *, statistically significant P value (P<0.05). CT, computed tomography; CTa, computed tomography attenuation value in the arterial phase; CTap, computed tomography arterial-phase enhancement value (CTap = CTa − CTmean); CTmax, maximum computed tomography attenuation value; CTmean, mean computed tomography attenuation value; CTmin, minimum computed tomography attenuation value; CTsd, standard deviation of computed tomography attenuation value; CTv, computed tomography attenuation value in the venous phase; CTvp, computed tomography venous-phase enhancement value (CTvp = CTv − CTmean); HU, Hounsfield unit; LD, long diameter; SD, short diameter.
Figure 4.
Representative CT images of benign and malignant pulmonary nodules. (A-D) Case 1: A 10-mm diameter pulmonary nodule was found in the left upper lobe. (A) The lung window showed regular nodule morphology without lobulation, spiculation, or other signs. (B) The mediastinal window showed that mean CT attenuation value of noncontrast scan was 57.52 HU. (C) The CT attenuation value in the arterial phase imaging was 78.45 HU. (D) The CT attenuation value in the venous phase imaging was 58.9 HU. The nodule was finally pathologically diagnosed as a pulmonary hamartoma. (E-H) Case 2: A 30-mm diameter pulmonary nodule in the left upper lobe. (E) The lung window showed the presence of lobulation, spiculation, and traction contact with the pleura. (F) The mediastinal window showed that the mean CT attenuation value of noncontrast scan was 63.29 HU. (G) The CT attenuation value in the arterial phase imaging was 56.06 HU. (H) The CT attenuation value in the venous phase imaging was 74.56 HU. The final pathological diagnosis of the solitary pulmonary nodule was lung adenocarcinoma. CT, computed tomography; HU, Hounsfield unit; SD, standard deviation.
Model construction and performance evaluation
A total of 34 candidate features were included in this study, consisting of 3 clinical features and 8 imaging features. Multivariable analysis identified independent risk factors, based on which four predictive models were constructed according to different CT phases. Model 1 was developed based on noncontrast CT images and incorporated 3 clinical features (age, surgical history, and hepatitis or liver cirrhosis) and 8 imaging features (interstitial lung disease, bronchiectasis, obstructive pneumonia, mGGO, lobulation, pleural retraction, CTmin, and CTmean). Models 2 incorporated arterial-phase CT parameters, model 3 incorporated venous-phase CT parameters, and model 4 incorporated noncontrast CT parameters together with dual-phase enhancement parameters. These models included identical independent predictors, comprising 3 clinical features (age, surgical history, and hepatitis or liver cirrhosis) and 7 imaging features (interstitial lung disease, bronchiectasis, obstructive pneumonia, mGGO, lobulation, pleural retraction, and CTmin). Although phase-specific enhancement parameters (CTa, CTap, CTv, and CTvp) were included as candidate variables in models 2–4, none of these enhancement-related variables remained statistically significant in the multivariable analysis and were therefore not retained in the final prediction model. As a result, the final versions of models 2–4 retained the same predictor set. Because the final predictor sets were identical, the diagnostic performance of Models 2–4 was highly consistent.
In the training cohort, all four models demonstrated good discriminative performance. Model 1 achieved the highest diagnostic efficacy, with an AUC of 0.928 (95% CI: 0.905–0.951), an accuracy of 86.8%, a sensitivity of 89.9%, and a specificity of 82.1%. Models 2–4 yielded identical AUC values of 0.914 (95% CI: 0.888–0.940), showing slightly inferior overall performance compared with model 1. Among them, models 2 and 4 exhibited higher specificity (both 87.0%) but lower sensitivity (both 80.8%), whereas model 3 showed higher sensitivity (87.8%) at the expense of lower specificity (77.7%).
In the test cohort, the diagnostic performance of all models further improved. Model 1 remained the best-performing model, with an AUC of 0.965 (95% CI: 0.940–0.989), an accuracy of 91.7%, a sensitivity of 92.3%, and a specificity of 90.9%. Models 2–4 again demonstrated identical AUC values of 0.963 (95% CI: 0.938–0.989). Models 2 and 4 maintained higher specificity (94.3%) but lower sensitivity (85.5%), while model 3 achieved a more balanced performance, with a sensitivity of 89.7% and a specificity of 90.9%.
In the external validation cohort, the performance of all models showed a slight decline compared with the training and test cohorts but remained clinically acceptable. Model 1 continued to demonstrate the best overall performance, with an AUC of 0.909 (95% CI: 0.870–0.949), a sensitivity of 85.2%, and a specificity of 74.0%. Models 2–4 showed identical AUC values of 0.904 (95% CI: 0.863–0.944). Among them, models 2 and 4 achieved higher specificity (84.9%) but lower sensitivity (75.5%), whereas model 3 exhibited the highest sensitivity (87.0%) with a relatively lower specificity (71.2%) (Table 3 and Figure 5A-5C). Calibration curves demonstrated good agreement between the predicted probabilities and observed outcomes in the training, test, and external validation cohorts, indicating satisfactory calibration performance for all four models (Figure 5D-5F).
Table 3. Comparison of the diagnostic efficacy between the four models.
| Dataset | Model type | AUC (95% CI) | Standard error | Accuracy (%) | Sensitivity (%) | Specificity (%) |
|---|---|---|---|---|---|---|
| Training | Model 1 | 0.928 (0.905–0.951) | 0.016 | 86.8 | 89.9 | 82.1 |
| Model 2 | 0.914 (0.888–0.940) | 0.017 | 83.2 | 80.8 | 87.0 | |
| Model 3 | 0.017 | 83.8 | 87.8 | 77.7 | ||
| Model 4 | 0.017 | 83.2 | 80.8 | 87.0 | ||
| Test | Model 1 | 0.965 (0.940–0.989) | 0.019 | 91.7 | 92.3 | 90.9 |
| Model 2 | 0.963 (0.938–0.989) | 0.022 | 89.3 | 85.5 | 94.3 | |
| Model 3 | 0.021 | 90.2 | 89.7 | 90.9 | ||
| Model 4 | 0.022 | 89.3 | 85.5 | 94.3 | ||
| Validation | Model 1 | 0.909 (0.870–0.949) | 0.029 | 80.9 | 85.2 | 74.0 |
| Model 2 | 0.904 (0.863–0.944) | 0.030 | 79.3 | 75.5 | 84.9 | |
| Model 3 | 0.029 | 80.9 | 87.0 | 71.2 | ||
| Model 4 | 0.030 | 79.3 | 75.5 | 84.9 |
Model 1: three clinical features (age, surgical history, and hepatitis or liver cirrhosis) and eight imaging features (interstitial lung disease, bronchiectasis, obstructive pneumonia, mixed ground-glass opacity, lobulation, pleural retraction, minimum CT attenuation value, mean CT attenuation value); Models 2–4: three clinical features (age, surgical history, and hepatitis or liver cirrhosis) and seven imaging features (interstitial lung disease, bronchiectasis, obstructive pneumonia, mixed ground-glass opacity, lobulation, pleural retraction, and minimum CT attenuation value). AUC, area under the curve; CI, confidence interval; CT, computed tomography.
Figure 5.
The ROC curves and calibration curves of the four models. (A-C) The ROC curves for the training, test, and validation cohorts. (D-F) The corresponding calibration curves. Model 1: three clinical features (age, surgical history, and hepatitis or liver cirrhosis) and eight imaging features (interstitial lung disease, bronchiectasis, obstructive pneumonia, mixed ground-glass opacity, lobulation, pleural retraction, minimum CT attenuation value, mean CT attenuation value); Models 2–4: three clinical features (age, surgical history, and hepatitis or liver cirrhosis) and seven imaging features (interstitial lung disease, bronchiectasis, obstructive pneumonia, mixed ground-glass opacity, lobulation, pleural retraction, and minimum CT attenuation value). AUC, area under the curve; CT, computed tomography; ROC, receiver operating characteristic.
Overall, model 1 consistently demonstrated the best and most stable diagnostic performance across all cohorts. Although models 2–4 shared identical AUC values, they exhibited different tradeoffs between sensitivity and specificity. DeLong tests were performed to compare the ROC curves of the four models. No statistically significant differences were observed among the models in the training, test, or external validation cohorts (all P>0.05), indicating that the inclusion of contrast-enhanced CT features did not significantly improve diagnostic performance. The optimal cutoff values for models 1–4 as determined by the maximum Youden index according to the ROC curves of the training cohort were 0.542, 0.675, 0.536, and 0.675, respectively. A confusion matrix was used to further evaluate the diagnostic performance of all models. and the accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated (Figure 6A-6L).
Figure 6.
Confusion matrix. (A-D) The confusion matrices of model 1 to model 4 in the training cohort. (E-H) The confusion matrices of the test cohort. (I-L) The confusion matrices of the validation cohort.
Stratified analysis of malignancy risk in pulmonary nodules based on CTap
Stratified analysis based on CTap was performed across the training, test, and external validation cohorts. An overall association between CTap levels and the probability of malignancy was observed in all three cohorts. Among the three stratification schemes, scheme 1 (<20, 20–40, and >40 HU) demonstrated the most consistent and stable pattern. In the test cohort, the malignancy rate increased progressively from 47.5% in the low-value group to 67.8% in the intermediate-value group and further to 79.2% in the high-value group. A similar successive increase was observed in the external validation cohort (54.8%, 69.2%, and 76.2%, respectively), indicating good reproducibility. In the training cohort, an increase from the low- to intermediate-value group was also observed, although a slight fluctuation was noted in the high-value group. In contrast, scheme 2 (<20, 20–60, and >60 HU) showed limited stability due to the relatively small sample size in the high-value subgroup, resulting in considerable variability in malignancy rates. Scheme 3 (<30, 30–50, and >50 HU) demonstrated a general increase with CTap levels; however, the gradient was less distinct compared to scheme 1 (Table 4 and Figure S1). Overall, these findings suggest that the probability of malignancy tends to increase with higher CTap levels, but this may plateau at higher enhancement levels (Figure S2).
Table 4. Proportion of malignant nodules and sample distribution for three CTap stratification schemes in the multicenter cohort.
| Group | CTap (HU) | Training, n (%) | Test, n (%) | Validation, n (%) | Total |
|---|---|---|---|---|---|
| 1 | <20 | 139 (55.6) | 58 (47.5) | 63 (54.8) | 115 |
| 20–40 | 90 (68.7) | 40 (67.8) | 36 (69.2) | 52 | |
| >40 | 57 (64.0) | 19 (79.2) | 16 (76.2) | 21 | |
| 2 | <20 | 139 (55.6) | 58 (47.5) | 63 (54.8) | 115 |
| 20–60 | 128 (69.2) | 55 (72.4) | 41 (68.3) | 60 | |
| >60 | 19 (54.3) | 4 (57.1) | 11 (84.6) | 13 | |
| 3 | <30 | 192 (58.7) | 80 (51.3) | 87 (58.4) | 149 |
| 30–50 | 60 (69.0) | 28 (77.8) | 15 (65.2) | 23 | |
| >50 | 34 (60.7) | 9 (69.2) | 13 (81.3) | 16 |
CTap, computed tomography arterial-phase enhancement value; HU, Hounsfield unit.
Discussion
In this study, we found that all four prediction models constructed based on clinical variables and three-phase CT data demonstrated high diagnostic performance in differentiating benign and malignant pulmonary nodules. The DeLong test revealed no statistically significant differences between the models (P>0.05), suggesting that the inclusion of contrast-enhanced CT features did not provide additional diagnostic value in this study. Among the four models, the nonenhanced CT-based model (model 1) achieved the highest AUC, indicating that fundamental imaging features derived from nonenhanced CT—such as nodule morphology, density, and size—remain important references for the initial assessment of pulmonary nodules. This finding is consistent with the conclusions reported by Snoeckx et al. regarding the importance of the morphological characteristics of pulmonary nodules (17). Our findings on the predictive value of conventional CT morphological features, such as lobulation and pleural retraction, are consistent with the observations reported by He et al. (18). In their study of small solid solitary pulmonary nodules (≤15 mm), they demonstrated that polygonal shape was a strong indicator of benign nodules, whereas lobulation, pleural retraction, and air bronchogram were more common in malignant nodules. Moreover, their results suggest that the number of discriminative CT features increases with nodule size, emphasizing the importance of stratifying nodules by size in the assessment of malignancy risk.
The comparable performance observed in the arterial-phase, venous-phase, and combined nonenhanced and enhanced models (models 2–4) may be attributed to the inclusion of the same clinical variables and CT features, indicating a limited incremental value of enhancement-related parameters. This finding differs from that of Zhang et al., who reported that a combined nonenhanced and contrast-enhanced CT model outperformed the nonenhanced model alone (16). In addition, Zhou et al. evaluated subtraction CT iodine maps for differentiating solitary pulmonary nodules. They reported that quantitative iodine values derived from arterial and venous subtraction maps could significantly improve the differentiation of benign, malignant, and inflammatory nodules when combined with morphological features (19). No incremental diagnostic value of contrast-enhanced CT was observed in our dataset, which may be due to several factors, including the multicenter and multivendor nature of the data, variability in scanning protocols, and interindividual differences in contrast administration. In contrast, nonenhanced CT may demonstrate better reproducibility across different scanners and acquisition settings, which could partly explain its superior performance.
Furthermore, we found that malignant pulmonary nodules exhibited lower enhancement CT values compared with benign nodules. This finding is inconsistent with the results reported by Yun et al., who reported significantly higher CTa and CTv values in malignant nodules than in benign ones (20). One possible explanation for this is the relatively high proportion of inflammatory nodules in the benign group included in our study. In general, inflammatory nodules tend to show marked enhancement on contrast-enhanced CT due to an abundant vascular supply or active inflammatory response. In contrast, certain malignant nodules, particularly those of early-stage lung cancers, may exhibit heterogeneous tumor angiogenesis, resulting in relatively lower enhancement compared with some benign lesions with a rich blood supply (21).
From a clinical perspective, our findings suggest that under certain conditions, nonenhanced CT alone may be sufficient for the preliminary differentiation of benign and malignant pulmonary nodules. This has important implications, including in reducing radiation exposure, lowering examination costs, and shortening scanning time. Moreover, it may facilitate the implementation of CT-based diagnostic strategies in primary healthcare settings. Therefore, the use of contrast-enhanced CT should be considered with greater caution, particularly in younger patients and those with impaired renal function.
It is widely believed that CT enhancement value can contribute substantially to the qualitative diagnosis of pulmonary nodules. Swensen et al. (22) proposed an enhancement threshold of benign and malignant pulmonary nodules of 20 HU, and Yamashita et al. increased this value to 20–60 HU (23). In the study by Yan et al., 118 of 167 cases of malignant pulmonary nodules had CT enhancement amplitude values between 20 and 60 HU (12), in line with the results reported by Yamashita et al. In our multicenter study, we performed a stratified analysis of arterial-phase CT attenuation values and demonstrated a stable association between CTap and the risk of malignancy in pulmonary nodules. Our findings indicate that relying primarily on a single enhancement threshold (e.g., 20–60 HU) to differentiate benign from malignant nodules has certain limitations. Specifically, the traditional 20 to 60 HU based stratification showed reduced stability in our dataset, mainly due to the limited sample size and increased variability within the high-value subgroup. By comparison, the stratification scheme comprising thresholds of <20, 20–40, and >40 HU exhibited superior stability and consistency in both the testing and external validation cohorts, providing a clearer and more clinically interpretable gradient of malignancy risk. These findings suggest that this alternative stratification approach may be more suitable for clinical risk assessment. Importantly, our results indicate that CTap should not be simply interpreted as a linear predictor or a fixed diagnostic cutoff. Instead, it may serve more effectively as a risk stratification tool, offering more nuanced information for clinical decision-making.
Our study involved certain limitations that should be acknowledged. First, we employed a multicenter retrospective design, and the generalizability and real-world applicability of the proposed model remain to be further validated. The stability and external applicability of the model in broader clinical settings may be questionable due to the differences in pathological composition between the cohorts, the potential selection bias inherent to retrospective studies, and the variability in CT scanners, acquisition parameters, and imaging protocols across centers. Second, although the CT feature analysis in this study was performed according to a standardized protocol, variability in CT data acquisition—including differences in scanners, acquisition parameters, and imaging protocols—may affect the accuracy of attenuation measurements and feature extraction. In addition, interobserver variability and differences in ROI selection may further weaken the association between imaging features and benign or malignant status, thereby reducing the overall predictive performance of the model. Notably, models 2–4 yielded identical predictors and similar performance, suggesting a high degree of redundancy among different contrast-enhanced phases. This finding implies that additional CT phases may not provide incremental diagnostic value, which could have practical implications for reducing radiation exposure and optimizing imaging protocols.
Future research should focus on several key directions. First, larger sample sizes and prospective multicenter studies are needed to further enhance the robustness and generalizability of the models. Second, the integration of computer-aided diagnosis techniques may enable automated and standardized extraction of CT features, thereby reducing observer-related variability and bias. Third, subgroup analyses based on pathological subtypes or nodule size should be conducted. Finally, machine learning and deep learning methods may be introduced to facilitate more comprehensive integration of multidimensional features and enhance the predictive capability of the model.
In addition, it should be noted that the prediction model developed in this study was based on a relatively homogeneous population of patients with pulmonary nodules, namely those without complex underlying lung diseases and with relatively simple baseline conditions. Therefore, its applicability in more complex clinical scenarios may be limited. For example, in patients with coexisting fibrotic interstitial lung diseases, background parenchymal abnormalities may significantly affect nodule detection, radiological feature assessment, and longitudinal evaluation. In such settings, conventional risk stratification models derived from nonfibrotic lungs may have limited discriminative value (24). Moreover, previous studies have shown that nodules arising in fibrotic lung regions, including small nodules and newly detected nodules, are more likely to be malignant than those in nonfibrotic lungs (25). Therefore, the applicability of the proposed model in these special populations remains to be further investigated.
Conclusions
The four predictive models constructed based on clinical data and three-phase CT imaging demonstrated consistently high diagnostic performance in differentiating benign from malignant pulmonary nodules, suggesting that CT enhancement features did not significantly improve model performance in and provided limited clinical added value. In addition, CTap stratification analysis indicated a potential association with the risk of malignancy in pulmonary nodules, possibly supporting its utility in risk stratification; however, further studies are needed to validate these findings.
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 study was approved by The First Affiliated Hospital of Bengbu Medical University (No. LWSL202300440), Anqing Municipal Hospital (No. 83230471), Tongde Hospital of Zhejiang Province (No. MR-33-24-041194), Hangzhou Hospital of Traditional Chinese Medicine (No. 2025KLL007), and Taizhou Municipal Hospital (No. LWYJ2023059). Informed consent was waived in this retrospective study.
Footnotes
Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0120/rc
Funding: This study was supported by the Zhejiang Traditional Chinese Medicine Administration (No. 2025ZL020, to J.W.), the Zhejiang Medical and Health Science and Technology Program (No. 2024KY1395, to H.C., and No. 2025KY1162, to L.D.), the Taizhou Science and Technology Plan Project (No. 25ywa33), the Cultivation Discipline Area Special Project (Quantum Medicine Project, 2024 Project) (No. 2024bypy017, to Z.X.), and the Longhu Talent Project of Bengbu Medical University (No. LH250302002, to Z.X.). The funder had no role in study design, data collection and analysis, manuscript writing, or the decision to submit the article for publication.
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0120/coif). All authors report grants from Zhejiang Traditional Chinese Medicine Administration, the Zhejiang Medical and Health Science and Technology Program, the Taizhou Science and Technology Plan Project, the Cultivation Discipline Area Special Project, and the Longhu Talent Project of Bengbu Medical University. The authors have no other conflicts of interest to declare.
Data Sharing Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0120/dss
References
- 1.Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. 10.3322/caac.21834 [DOI] [PubMed] [Google Scholar]
- 2.Han B, Zheng R, Zeng H, Wang S, Sun K, Chen R, Li L, Wei W, He J. Cancer incidence and mortality in China, 2022. J Natl Cancer Cent 2024;4:47-53. 10.1016/j.jncc.2024.01.006 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Du W, He B, Luo X, Chen M. Diagnostic Value of Artificial Intelligence Based on CT Image in Benign and Malignant Pulmonary Nodules. J Oncol 2022;2022:5818423. 10.1155/2022/5818423 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Mankidy BJ, Mohammad G, Trinh K, Ayyappan AP, Huang Q, Bujarski S, Jafferji MS, Ghanta R, Hanania AN, Lazarus DR. High risk lung nodule: A multidisciplinary approach to diagnosis and management. Respir Med 2023;214:107277. 10.1016/j.rmed.2023.107277 [DOI] [PubMed] [Google Scholar]
- 5.Anderson IJ, Davis AM. Incidental Pulmonary Nodules Detected on CT Images. JAMA 2018;320:2260-1. 10.1001/jama.2018.16336 [DOI] [PubMed] [Google Scholar]
- 6.Mazzone PJ, Lam L. Evaluating the Patient With a Pulmonary Nodule: A Review. JAMA 2022;327:264-73. 10.1001/jama.2021.24287 [DOI] [PubMed] [Google Scholar]
- 7.Sands J, Tammemägi MC, Couraud S, Baldwin DR, Borondy-Kitts A, Yankelevitz D, Lewis J, Grannis F, Kauczor HU, von Stackelberg O, Sequist L, Pastorino U, McKee B. Lung Screening Benefits and Challenges: A Review of The Data and Outline for Implementation. J Thorac Oncol 2021;16:37-53. 10.1016/j.jtho.2020.10.127 [DOI] [PubMed] [Google Scholar]
- 8.Brodersen J, Voss T, Martiny F, Siersma V, Barratt A, Heleno B. Overdiagnosis of lung cancer with low-dose computed tomography screening: meta-analysis of the randomised clinical trials. Breathe (Sheff) 2020;16:200013. 10.1183/20734735.0013-2020 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Duan X, Ouyang Z, Bao S, Yang L, Deng A, Zheng G, Zhu Y, Li G, Chu J, Liao C. Factors associated with overdiagnosis of benign pulmonary nodules as malignancy: a retrospective cohort study. BMC Pulm Med 2023;23:454. 10.1186/s12890-023-02727-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Zheng C, Wang H, Liu Q, Han D, Xin Y, Lu W, Yan Z. Application effect of low-dose spiral CT on pulmonary nodules and its diagnostic value for benign and malignant nodules. Am J Transl Res 2023;15:256-63. [PMC free article] [PubMed] [Google Scholar]
- 11.Wu H, Zhang X, Zhong Z. Exploration of CT-based discrimination and diagnosis of various pathological types of ground glass nodules in the lungs. BMC Med Imaging 2025;25:119. 10.1186/s12880-025-01653-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Yan G, Li H, Fan X, Deng J, Yan J, Qiao F, Yan G, Liu T, Chen J, Wang L, Yang Y, Li Y, Zhao L, Bhetuwal A, McClure MA, Li N, Peng C. Multimodality CT imaging contributes to improving the diagnostic accuracy of solitary pulmonary nodules: a multi-institutional and prospective study. Radiol Oncol 2023;57:20-34. 10.2478/raon-2023-0008 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.EL-Metwally D , EL-Bakry MAH, EL-Adalany MA, Adel S, Mohsen M. The role of dynamic contrast-enhanced CT in characterization of solitary solid pulmonary nodules. Egypt J Radiol Nucl Med 2022;53:231. [Google Scholar]
- 14.Zhao L, Lv Y, Zhou Y, Wu A, Yang D, Shi H, Wang J, Lin M. Establishing predictive models for malignant and inflammatory pulmonary nodules using clinical data and CT imaging features. Quant Imaging Med Surg 2025;15:2957-70. 10.21037/qims-24-2338 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Nissar A, Mir AH. Predictive radiomics based ensemble machine learning approach in CT lung nodule diagnosis. J Egypt Natl Canc Inst 2025;37:68. 10.1186/s43046-025-00326-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Zhang W, Cui X, Wang J, Cui S, Yang J, Meng J, Zhu W, Li Z, Niu J. The study of plain CT combined with contrast-enhanced CT-based models in predicting malignancy of solitary solid pulmonary nodules. Sci Rep 2024;14:21871. 10.1038/s41598-024-72592-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Snoeckx A, Reyntiens P, Desbuquoit D, Spinhoven MJ, Van Schil PE, van Meerbeeck JP, Parizel PM. Evaluation of the solitary pulmonary nodule: size matters, but do not ignore the power of morphology. Insights Imaging 2018;9:73-86. 10.1007/s13244-017-0581-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.He XQ, Huang XT, Luo TY, Liu X, Li Q. The differential computed tomography features between small benign and malignant solid solitary pulmonary nodules with different sizes. Quant Imaging Med Surg 2024;14:1348-58. 10.21037/qims-23-995 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Zhou W, Qu S, Guo ZZ, Qiao WD, Sun TX, Xu SY, Zhao XD, Lu X, Liu XY. The capacity of subtraction CT iodine maps to improve the differentiation of solitary pulmonary nodules. Quant Imaging Med Surg 2025;15:11336-50. 10.21037/qims-2025-1083 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Yun P, Zhan Z, Wu Z, Rao Y, Sun K, Gong L. The added value of quantitative contrast-enhanced CT parameters in distinguishing malignant from benign solid pulmonary nodules. BMC Med Imaging 2025;25:294. 10.1186/s12880-025-01832-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Zhang M, Kono M. Solitary pulmonary nodules: evaluation of blood flow patterns with dynamic CT. Radiology 1997;205:471-8. 10.1148/radiology.205.2.9356631 [DOI] [PubMed] [Google Scholar]
- 22.Swensen SJ, Brown LR, Colby TV, Weaver AL. Pulmonary nodules: CT evaluation of enhancement with iodinated contrast material. Radiology 1995;194:393-8. 10.1148/radiology.194.2.7824716 [DOI] [PubMed] [Google Scholar]
- 23.Yamashita K, Matsunobe S, Tsuda T, Nemoto T, Matsumoto K, Miki H, Konishi J. Solitary pulmonary nodule: preliminary study of evaluation with incremental dynamic CT. Radiology 1995;194:399-405. 10.1148/radiology.194.2.7824717 [DOI] [PubMed] [Google Scholar]
- 24.Pianigiani T, Cameli P, Di Lorenzo M, Bergantini L, Bargagli E. Longitudinal Multimodal Assessment of Pulmonary Nodules in Fibrosing Interstitial Lung Diseases: A Retrospective Study. Lung 2026;204:11. 10.1007/s00408-026-00877-z [DOI] [PubMed] [Google Scholar]
- 25.Salvatore MM, Liu Y, Peng B, Hsu HY, Saqi A, Tsai WY, Leu CS, Jambawalikar S. Comparison of lung cancer occurring in fibrotic versus non-fibrotic lung on chest CT. J Transl Med 2024;22:67. 10.1186/s12967-023-04645-y [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
The article’s supplementary files as
Data Availability Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0120/dss