Abstract
Background
Lung cancer remains one of the leading causes of cancer-related deaths worldwide. This study utilized clinical risk factors along with intratumoral radiomics, peritumoral radiomics, and intratumoral subregional features extracted from computed tomography (CT) lung-window images for individual and integrated modeling to classify solid pulmonary nodules and identify the optimal model, thereby improving diagnostic accuracy while minimizing unnecessary invasive procedures.
Methods
CT images of 230 pathologically confirmed solitary solid pulmonary nodules were retrospectively collected from two hospitals. Among the 167 patients from the first hospital, 20% (n=34) served as the test set, while the remaining 80% (n=133) were used as the training and development set for 5-fold cross-validation, while data from the second hospital (n=63) served as an external test set. Intratumoral and peritumoral regions of interest (ROIs) were delineated on lung window images, and relevant radiomics features were extracted. Multiple machine learning algorithms—including Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Linear Support Vector Classifier (Linear SVC) etc.—were employed to construct predictive models for distinguishing benign from malignant solid pulmonary nodules.
Results
A triple-feature model (intratumoral, peritumoral, clinical) achieved superior diagnostic performance [area under the receiver operating characteristic curve (AUC): training 0.932, 95% confidence interval (CI): 0.897–0.960; test 0.833, 95% CI: 0.773–0.890; external test 0.741, 95% CI: 0.618–0.864] with high sensitivity/specificity. The intratumoral-peritumoral dual-modality model showed optimal cross-center robustness external test, AUC =0.808 (95% CI: 0.700–0.922). Habitat imaging revealed heterogeneity, AUC =0.750 (95% CI: 0.676–0.825). Decision curve analysis confirmed the triple-model’s clinical utility. SHAP identified age, gender, and key radiomics (e.g., gradient_firstorder_Skewness_Intra) as top predictors. Multi-center test confirmed generalizability, positioning this integrated framework as a robust tool to reduce invasive procedures in pulmonary nodule management.
Conclusions
The multi-combination models developed in this study enhance the diagnostic accuracy for distinguishing benign from malignant solid pulmonary nodules, with the triple-feature model demonstrating the highest diagnostic performance. This approach has the potential to spare patients from unnecessary invasive procedures and strengthen clinical decision-making in the management of pulmonary nodules.
Keywords: Lung, radiomics, machine learning
Highlight box.
Key findings
• A triple-feature model combining intratumoral radiomics, peritumoral radiomics, and clinical data achieved the best diagnostic performance [area under the receiver operating characteristic curve (AUC): 0.932 in training, 0.833 in test, 0.741 in external test] for distinguishing benign from malignant solid pulmonary nodules.
• The intratumoral-peritumoral dual-modality model demonstrated superior cross-center robustness (external test AUC =0.808).
• Habitat (subregional) imaging captured spatial heterogeneity but did not outperform conventional radiomics models in diagnostic accuracy.
What is known and what is new?
• Radiomics based on intratumoral and peritumoral regions has been increasingly used to classify pulmonary nodules, yet the added value of intratumoral habitat (subregional) analysis and its integration with clinical factors in multicenter settings remains underexplored.
• This study systematically compares multiple feature combinations, validates models across two independent centers, and demonstrates that while habitat imaging provides supplementary information, a combined intratumoral-peritumoral-clinical model offers the highest diagnostic accuracy in internal validation, though its generalizability is affected by cross-center parameter variations.
What is the implication, and what should change now?
• This study supports the integration of peritumoral radiomics and clinical factors into pulmonary nodule assessment frameworks to improve diagnostic precision.
• Future research should prioritize prospective standardization of computed tomography acquisition protocols and develop domain adaptation methods to enhance model generalizability across institutions.
Introduction
Lung cancer remains the leading cause of cancer-related deaths globally, with projected fatalities in 2022 surpassing the combined mortality of colorectal, breast, and prostate cancers (1,2). According to the International Agency for Research on Cancer, 2020 witnessed approximately 2.2 million new cases and 1.8 million deaths worldwide, with developing countries facing persistently increasing epidemiological burdens (3-5). In the field of early lung cancer diagnosis, pulmonary nodules serve as key imaging biomarkers, of which their density characteristics hold significant clinical relevance. Based on differences in computed tomography (CT) attenuation values, pulmonary nodules are classified into two categories: ground-glass nodules (GGNs) and solid nodules (6). Notably, solid nodules in lung cancer exhibit stronger biological aggressiveness compared to GGNs: shorter doubling times, higher incidence of lymph node metastasis, and significantly reduced 5-year survival rates. These factors underscore the clinical imperative for precise diagnosis of solid nodules to guide timely interventions (7-11).
Current imaging-based evaluation of solid pulmonary nodules faces dual challenges. Firstly, diagnostic efficacy varies significantly with nodule size. Studies indicate that the malignancy probability of nodules measuring 8–30 mm in diameter spans a wide range (1–70%), necessitating reliance on the combined analysis of risk factors and morphological characteristics (12). Secondly, morphological features exhibit limited specificity for malignancy. Although signs such as pleural indentation and spiculation are widely used for differentiation, overlapping features frequently occur in 8–30 mm nodules, leading to markedly reduced accuracy of conventional CT diagnosis (8,13,14).
This diagnostic uncertainty directly impacts the quality of clinical decision-making. Data from the National Lung Screening Trial revealed that 40% of nodules detected by low-dose CT were pathologically benign post-resection, while studies in Chinese populations reported benign rates of 25–35%. Such gaps in diagnostic performance result in healthcare resource wastage and patient psychological distress, underscoring the limitations of conventional imaging interpretation in distinguishing benign from malignant nodules (12,15,16).
In this context, radiomics—a technology leveraging standardized feature extraction and machine learning modeling—offers a novel pathway to transcend human visual interpretation limits (17). Research by Warkentin et al. (18) demonstrated that radiomics-based models exhibit superior performance in the differentiation of solid pulmonary nodules, outperforming the established Pan-Canadian Early Lung Cancer Detection model [area under the receiver operating characteristic curve (AUC) 0.868] in nodule assessment. However, traditional radiomics studies often assume homogeneous intratumoral heterogeneity and generate averaged quantitative features across the entire tumor, thereby neglecting spatial heterogeneity within the tumor microenvironment (19,20). The inability to resolve this functional heterogeneity may represent a critical bottleneck limiting the clinical translation of existing models (21,22). To address this bottleneck, habitat analysis offers a novel approach by partitioning the tumor into distinct functional subregions, enabling the resolution of spatial heterogeneity. Researchers may delineate functional subregions using imaging intensity gradients and textural heterogeneity, subsequently leveraging tumor subregional features to guide diagnosis and even inform precision treatment strategies (23,24). While habitat analysis holds the potential to inform diagnosis and even precision treatment strategies, its comparative value against conventional intra- and peritumoral radiomics in the context of solid pulmonary nodules remains to be fully established, particularly given challenges such as imaging resolution and feature stability.
To address the challenge of distinguishing between benign and malignant solid pulmonary nodules in clinical practice, this study employs multiple machine learning methods to develop predictive models for distinguishing benign and malignant solid pulmonary nodules based on clinical risk factors and CT radiomics features. The goal is to establish a more accurate benign-malignant prediction framework, providing robust evidence for clinical decision-making. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1869/rc).
Methods
Study population
The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the respective Ethics Committees of Shaoxing People’s Hospital (No. 2021-K-Y-021-01) and Shaoxing Second Hospital (No. 2024029001). Informed consent was waived in this retrospective study. From January 2020 to December 2024, a total of 292 lung solid nodule patients who underwent partial pneumonectomy or lung needle biopsy were included from Shaoxing People’s Hospital and Shaoxing Second Hospital, with pathological confirmation of either lung adenocarcinoma or granulomatous lung disease. Based on the inclusion and exclusion criteria outlined in Figure 1, 230 patients were enrolled in the final analysis. The patient distribution across the institutions was as follows: 167 patients from Database 1 (Shaoxing People’s Hospital) and 63 patients from Database 2 (Shaoxing Second Hospital). In Database 1, 20% (n=34) of the 167 patients served as the test set, while the remaining 80% (n=133) were used as the training set for 5-fold cross-validation. The 63 patients from Database 2 were designated as the external test set.
Figure 1.
From January 2020 to December 2024, patients who underwent partial pneumonectomy or lung needle biopsy were included from two distinct medical institutions. The inclusion and exclusion criteria for enrolled patients are shown above. CT, computed tomography; ROIs, regions of interest.
Clinical and pathological data were collected and reviewed, including: (I) demographic and clinical characteristics: sex, age, smoking history, and prior tumor history; (II) CT imaging features: lesion location, maximum tumor diameter, lobulation, spiculation, vascular convergence, pleural traction, regularity of morphology, clarity of boundaries, homogeneity of density, vascular passing-through sign, and calcification; (III) pathological findings: histopathological diagnosis retrieved from pathology reports. All pathological diagnoses were reported by pathologists with at least 3 years of experience and reviewed by experts with over 10 years of experience to ensure accuracy.
CT image acquisition parameters and preprocessing
The data for this retrospective study were obtained from two medical centers. The details are as follows: Database 1 (training and test sets): a 64-slice CT scanner (Brilliance 64, Philips Healthcare, the Netherlands) was used. Scanning parameters: tube voltage 120 kV, tube current 200–250 mAs, rotation time 0.75 s, reconstruction matrix 512×512. The original scan slice thickness and spacing were 1.0 mm, and post-processing thin-slice reconstruction was performed at 0.67 mm. Database 2 (external test set): A 64-slice CT scanner (SOMATOM Definition AS, Siemens Healthineers, Germany) was used. Scanning parameters: tube voltage 120 kV, tube current 200–300 mAs, rotation time 0.75 s, reconstruction matrix 512×512. The original scan slice thickness and spacing were 5.0 mm, and post-processing thin-slice reconstruction was performed at 0.6–1.5 mm. To minimize feature variability caused by differences in scanning devices and protocols, all images included in the analysis (all of which were thin-slice reconstruction data provided by each center) underwent a uniform standardized preprocessing pipeline before feature extraction: firstly, resampling: all images were resampled to an isotropic resolution (1 mm × 1 mm × 1 mm) using B-spline interpolation to eliminate variations in original slice thickness and pixel spacing. Secondly, intensity normalization: CT values were first truncated using a standard lung window [window width 1,500 Hounsfield units (HU), window level –600 HU] and then linearly normalized to the [0, 1] range. This standardized workflow was designed to improve the reproducibility and comparability of radiomics features across different technical conditions and served as the foundation for subsequent cross-center analysis.
Image segmentation
All patients underwent CT examination. Region of interest (ROI) segmentation was performed by two experienced radiologists using the Deepwise multimodal research platform version 2.6.4 (https://keyan.deepwise.com). When there is a significant dispute over the definition between two radiologists, the final ROI area is determined by a third senior doctor. Use a research platform to automatically expand ROI 2 mm outward to obtain a circular area around the tumor (25). Use SLIC algorithm to perform superpixel segmentation on CT images, attempting different segmentation numbers (i ranging from 2 to 4), and ultimately generating 3 segmentation subregion results. Calculate the average intensity of each region, and reassign labels based on the average intensity sorting results. To ensure the stability of the delineation, 20 cases were randomly selected 2 weeks after the completion of the delineation, and then two radiologists conducted the delineation again. Intra-rater agreement and inter-rater agreement were assessed using the intraclass correlation coefficient (ICC) based on the 95% confidence interval (CI) of the ICC estimate, with a threshold of ICC ≥0.8, to evaluate the consistency of measurement results.
Radiomics features extraction and selection
The radiomics analysis workflow for this study is illustrated in Figure 2, encompassing three primary steps: ROI segmentation, feature extraction and selection, and model construction and evaluation. Radiomics feature extraction was conducted using the Deepwise multimodal research platform, incorporating shape features, first-order features, and texture features. The texture features included metrics derived from the Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Gray Level Difference Matrix (GLDM), and Neighborhood Gray Tone Difference Matrix (NGTDM). To enhance the original features, preprocessing techniques such as Wavelet, Exponential, Gradient, Local Binary Patterns in 2D (LBP2D), Local Binary Patterns in 3D (LBP3D), Logarithm, Square, Square Root, and Laplacian of Gaussian (LoG) filtering were applied. Normalization was disabled (normalize: False), and a padding distance of 10 was applied to accommodate large sigma values in LoG filtering. The minimum ROI dimensions were set to 2, and image discretization was performed with a bin width of 25. For first-order features, a voxel array shift of 1,000 was applied to address negative Hounsfield Units. Additionally, a default label value of 1 was used, and pre-cropping was enabled (preCrop: true) to ensure accurate feature extraction. A total of 2,153 features were extracted from the CT images. After ICC analysis, 1,839 intratumoral features, 1,673 peritumoral features, 1,596 intratumoral subregion features were retained.
Figure 2.
The radiomics analysis workflow for this study, encompassing three primary steps: ROI segmentation, feature extraction and selection, and model construction and evaluation. GLCM, Gray Level Co-occurrence Matrix; GLDM, Gray Level Difference Matrix; GLRLM, Gray Level Run Length Matrix; GLSZM, Gray Level Size Zone Matrix; ICC, intraclass correlation coefficient; NGTDM, Neighborhood Gray Tone Difference Matrix; ROI, region of interest.
Firstly, we conducted feature Pearson correlation analysis to reduce redundancy with a threshold value of 0.8, removing one of the features if the linear correlation coefficient between any two independent variables in the training set exceeded the specified threshold; subsequently, to determine the optimal parameters for radiomics feature selection based on the analysis of variance (ANOVA F-value), a grid search approach was adopted, specifically selecting the top 20 to 40 features in increments of 5 for ANOVA F-value-based feature selection to identify the optimal number of features for each method, with all radiomics feature selection processes performed using the Deepwise Multimodal Research Platform version 2.5.1 (https://keyan.deepwise.com, Hangzhou Deepwise League of PHD Technology Co., Ltd., Hangzhou, China). The number of final modeling features for each model is as follows: intratumoral + peritumoral + clinical (26 features), intratumoral + clinical (64 features), intratumoral subregion + clinical (30 features), peritumoral + clinical (70 features), intratumoral + peritumoral (80 features), intratumoral subregion + peritumoral (30 features), intratumoral (30 features), intratumoral subregion (30 features), peritumoral (30 features), and clinical (19 features).
Development, performance, and validation of multiple models
After refining the features, we tried 10 data combinations to develop predictive models, including intratumoral + peritumoral + clinical, intratumoral + clinical, intratumoral subregion + clinical, peritumoral + clinical, intratumoral + peritumoral, intratumoral subregion + peritumoral, intratumoral, intratumoral subregion, peritumoral, clinical, etc. To avoid overfitting and fairly evaluate the performance of the model, a nested five-fold cross-validation strategy was adopted, with four folds used as the training set and one-fold used as the test set in the outer loop. The above steps of feature screening were carried out separately in each cycle. In the inner cycle, the optimum hyper-parameters were selected by using fivefold cross-validation for the training set. Logistic Regression (LR), Support Vector Machine (SVM), Linear Support Vector Classifier (Linear SVC), Decision Tree, Random Forest, Ada Boost, Gradient Boosting, XG Boost, Bernoulli NB, Gaussian NB, K Nearest Neighbors, Linear Discriminant Analysis, SGD, Multilayer Perceptron Classifier algorithm was used to develop the predictive model. A systematic hyperparameter search was conducted to optimize the classifiers. The search focused on the penalty, fit intercept, kernel, dual, etc. The search process was repeated 10 times, with each iteration evaluating using cross-validation. The performance of the model was evaluated on the test set and external test set. After a round of fivefold cross-validation, all the data were used for performance evaluation. AUC, sensitivity, specificity, accuracy and other indicators were adopted for model performance evaluation. The efficacy and generalization capability of the predictive models were assessed by plotting the receiver operating characteristic (ROC) curves, calibration curves, and the decision curve analysis (DCA). Additionally, SHapley Additive exPlanations (SHAP) analysis enhanced transparency and clinical trust by clearly elucidating the contribution of each feature to the model’s predictions, the work flow of the study in Figure 2. The development, performance evaluation, and validation of multiple models were conducted using the Deepwise Multimodal Research Platform version 2.5.1 (https://keyan.deepwise.com, Hangzhou Deepwise League of PHD Technology Co., Ltd., Hangzhou, China).
Statistical analysis
Statistical analyses were performed using SPSS software (version 26.0), R program software (version 4.3.3), and Python software (version 3.7.13), with a significance level set at P<0.05. Continuous variables were expressed as mean ± standard deviation or quartiles, and categorical variables were presented as frequencies and percentages. Normality of the data was assessed using the Shapiro-Wilk test. Based on the results of this test, the choice of statistical methods was determined as follows: for independent continuous variables, a two-sample independent t-test was used if the data followed a normal distribution. Wilcoxon rank sum test is typically used when comparing two independent groups of ordinal or continuous data that do not meet the assumptions of normality. It is a non-parametric test that assesses whether the distributions of the two groups are significantly different from each other, often applied in scenarios where the data is skewed or the sample size is small. Chi-squared test is used to determine whether there is a significant association between two categorical variables in a contingency table. Yates’ Corrected Chi-squared test is a modified version of the Chi-squared test that adjusts for continuity, making it more appropriate for small sample sizes, particularly in 2×2 contingency tables. It is used when the expected frequencies are low, helping to reduce the overestimation of statistical significance that can occur with the standard Chi-squared test in such cases.
Results
Patient characteristics
Finally, a total of 167 patients were included in the internal test set, comprising 70 cases of granulomatous lung disease and 97 cases of lung adenocarcinoma. For external test, 63 patients were enrolled, including 27 patients with granulomatous lung disease and 36 patients with lung adenocarcinoma. In 2 cases of lung adenocarcinoma, discontinuous lesions were observed across imaging slices, necessitating the delineation of two separate ROIs. The baseline characteristics of both the internal and external sets are summarized in Tables 1,2, respectively. In univariate analysis, age, maximum tumor diameter and lobulation sign in the internal set, along with Gender, density homogeneous and clarity of boundaries in the external set, were significant variables for differentiating granulomatous lung disease from lung adenocarcinoma in solid pulmonary nodules (each P<0.05, Tables 1,2).
Table 1. Baseline characteristics of patients level in internal data center.
| Variable | Granulomatous lung disease (N=69) | Lung adenocarcinoma (N=98) | P |
|---|---|---|---|
| Gender | 0.17† | ||
| Female | 25 (36.2) | 46 (46.9) | |
| Male | 44 (63.8) | 52 (53.1) | |
| Age (years) | 57.246±11.718 | 62.510±10.819 | 0.003**‡ |
| Maximum diameter of tumor on lung window (mm) | 14.029±5.216 | 16.214±5.933 | 0.012*‡ |
| Previous tumor history | 0.74† | ||
| No | 63 (91.3) | 88 (89.8) | |
| Yes | 6 (8.7) | 10 (10.2) | |
| Smoking | 0.42† | ||
| No | 46 (66.7) | 71 (72.4) | |
| Yes | 23 (33.3) | 27 (27.6) | |
| CEA (ng/mL) | 2.716±1.927 | 3.199±2.352 | 0.16‡ |
| CA125 (U/mL) | 11.427±5.460 | 22.794±73.171 | 0.20‡ |
| Squamous cell carcinoma antigen (ng/mL) | 0.903±0.554 | 0.829±0.358 | 0.30‡ |
| Neuron-specific enolase (ng/mL) | 3.550 (2.790–4.640) | 3.255 (2.645–3.935) | 0.055§ |
| Location | 0.85† | ||
| Upper lobe of right lung | 22 (31.9) | 34 (34.7) | |
| Lower lobe of right lung | 7 (10.1) | 8 (8.2) | |
| Middle lobe of right lung | 11 (15.9) | 21 (21.4) | |
| Upper lobe of left lung | 15 (21.7) | 19 (19.4) | |
| Lower lobe of left lung | 14 (20.3) | 16 (16.3) | |
| Lobulation | 0.02*† | ||
| No | 15 (21.7) | 9 (9.2) | |
| Yes | 54 (78.3) | 89 (90.8) | |
| Spiculation | 0.86† | ||
| No | 22 (31.9) | 30 (30.6) | |
| Yes | 47 (68.1) | 68 (69.4) | |
| Pleural traction | 0.45† | ||
| No | 22 (31.9) | 26 (26.5) | |
| Yes | 47 (68.1) | 72 (73.5) | |
| Vascular passing-through sign | 0.60¶ | ||
| No | 1 (1.4) | 4 (4.1) | |
| Yes | 68 (98.6) | 94 (95.9) | |
| Vascular convergence | >0.99¶ | ||
| No | 67 (97.1) | 94 (95.9) | |
| Yes | 2 (2.9) | 4 (4.1) | |
| Calcification | 0.19¶ | ||
| No | 65 (94.2) | 97 (99.0) | |
| Yes | 4 (5.8) | 1 (1.0) | |
| Is the density homogeneous | 0.20† | ||
| No | 29 (42.0) | 51 (52.0) | |
| Yes | 40 (58.0) | 47 (48.0) | |
| Is the boundary clear | 0.99† | ||
| No | 24 (34.8) | 34 (34.7) | |
| Yes | 45 (65.2) | 64 (65.3) | |
| Is the shape regular | 0.25† | ||
| No | 52 (75.4) | 81 (82.7) | |
| Yes | 17 (24.6) | 17 (17.3) |
Data are presented as n (%), mean ± standard deviation or median (interquartile range). *, P<0.05; **, P<0.01; †, Chi-squared test; ‡, t-test; §, Wilcoxon rank sum test; ¶, Yates’ Corrected Chi-squared test. CA125, cancer antigen 125; CEA, carcinoembryonic antigen.
Table 2. Baseline characteristics of patients level in external data center.
| Variable | Granulomatous lung disease (N=27) | Lung adenocarcinoma (N=38) | P |
|---|---|---|---|
| Gender | 0.003**† | ||
| Female | 9 (33.3) | 27 (71.1) | |
| Male | 18 (66.7) | 11 (28.9) | |
| Age (years) | 60.444±13.568 | 62.158±9.440 | 0.55‡ |
| Maximum diameter of tumor on lung window (mm) | 17.296±5.941 | 16.184±5.849 | 0.46‡ |
| Previous tumor history | 0.76§ | ||
| No | 25 (92.6) | 37 (97.4) | |
| Yes | 2 (7.4) | 1 (2.6) | |
| Smoking | 0.58§ | ||
| No | 22 (81.5) | 34 (89.5) | |
| Yes | 5 (18.5) | 4 (10.5) | |
| CEA (ng/mL) | 2.764±1.540 | 3.229±1.393 | 0.21‡ |
| CA125 (U/mL) | 13.200 (8.470–16.250) | 9.550 (8.700–12.550) | 0.11¶ |
| Squamous cell carcinoma antigen (ng/mL) | 0.873±0.346 | 0.772±0.281 | 0.20‡ |
| Neuron-specific enolase (ng/mL) | 2.886±2.095 | 2.798±1.306 | 0.84‡ |
| Location | 0.34† | ||
| Upper lobe of right lung | 6 (22.2) | 16 (42.1) | |
| Lower lobe of right lung | 5 (18.5) | 6 (15.8) | |
| Middle lobe of right lung | 3 (11.1) | 4 (10.5) | |
| Upper lobe of left lung | 8 (29.6) | 10 (26.3) | |
| Lower lobe of left lung | 5 (18.5) | 2 (5.3) | |
| Lobulation | 0.054§ | ||
| No | 4 (14.8) | 0 (0.0) | |
| Yes | 23 (85.2) | 38 (100.0) | |
| Spiculation | 0.053† | ||
| No | 15 (55.6) | 12 (31.6) | |
| Yes | 12 (44.4) | 26 (68.4) | |
| Pleural traction | 0.87† | ||
| No | 8 (29.6) | 12 (31.6) | |
| Yes | 19 (70.4) | 26 (68.4) | |
| Vascular passing-through sign | 0.86§ | ||
| No | 1 (3.7) | 0 (0.0) | |
| Yes | 26 (96.3) | 38 (100.0) | |
| Vascular convergence | >0.99§ | ||
| No | 26 (96.3) | 37 (97.4) | |
| Yes | 1 (3.7) | 1 (2.6) | |
| Calcification | 0.86§ | ||
| No | 26 (96.3) | 38 (100.0) | |
| Yes | 1 (3.7) | 0 (0.0) | |
| Is the density homogeneous | 0.04*§ | ||
| No | 21 (77.8) | 37 (97.4) | |
| Yes | 6 (22.2) | 1 (2.6) | |
| Is the boundary clear | 0.001**† | ||
| No | 13 (48.1) | 33 (86.8) | |
| Yes | 14 (51.9) | 5 (13.2) | |
| Is the shape regular | 0.33§ | ||
| No | 25 (92.6) | 38 (100.0) | |
| Yes | 2 (7.4) | 0 (0.0) |
Data are presented as n (%), mean ± standard deviation or median (interquartile range). *, P<0.05; **, P<0.01; †, Chi-squared test; ‡, t-test; §, Yates’ Corrected Chi-squared test; ¶, Wilcoxon rank sum test. CA125, cancer antigen 125; CEA, carcinoembryonic antigen.
Performance of nine models
Based on the imaging images of the intratumoral region, 2 mm peritumoral region, and intratumoral subregion, we extracted 2,153 features from these images, and then combined each set of features with clinical data for feature selection. Constructing a composite radiomics model through least absolute shrinkage and selection operator (LASSO) dimensionality reduction, including combinations of intratumoral + peritumoral + clinical, intratumoral + clinical, intratumoral subregion + clinical, peritumoral + clinical, intratumoral + peritumoral, intratumoral subregion + peritumoral, intratumoral, intratumoral subregion, peritumoral, clinical, etc. Quantitatively evaluate the differences between Lung adenocarcinoma and Granulomatous lung disease. The performance of these models on the training set, test set and external test set is shown in Figure 3 and Table 3.
Figure 3.
The performance of these models on the training set, test set and external test set. AUC, area under the curve; ROC, receiver operating characteristic.
Table 3. The performance of model.
| Model | Train cohort | Test cohort | External test cohort | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AUC (95% CI) | Accuracy | Sensitivity | Specificity | AUC (95% CI) | Accuracy | Sensitivity | Specificity | AUC (95% CI) |
Accuracy | Sensitivity | Specificity | |||
| Intratumoral + peritumoral + clinical | 0.932 (0.897–0.960) |
0.893 | 0.847 | 0.857 | 0.833 (0.773–0.890) |
0.750 | 0.735 | 0.771 | 0.741 (0.618–0.864) |
0.646 | 0.711 | 0.556 | ||
| Intratumoral + clinical | 0.911 (0.869–0.953) |
0.833 | 0.908 | 0.729 | 0.759 (0.689–0.829) |
0.691 | 0.776 | 0.571 | 0.606 (0.468–0.744) |
0.569 | 0.474 | 0.704 | ||
| Intratumoral subregion + clinical | 0.776 (0.704–0.848) |
0.705 | 0.804 | 0.565 | 0.745 (0.669–0.821) |
0.716 | 0.804 | 0.551 | 0.593 (0.506–0.679) |
0.557 | 0.561 | 0.551 | ||
| Peritumoral + clinical | 0.897 (0.852–0.941) |
0.798 | 0.918 | 0.629 | 0.737 (0.663–0.810) |
0.673 | 0.847 | 0.429 | 0.576 (0.430–0.723) |
0.477 | 0.290 | 0.741 | ||
| Intratumoral + peritumoral | 0.858 (0.808–0.907) |
0.781 | 0.860 | 0.670 | 0.789 (0.729–0.848) |
0.743 | 0.853 | 0.588 | 0.808 (0.700–0.922) |
0.739 | 0.868 | 0.556 | ||
| Intratumoral subregion + peritumoral | 0.767 (0.694–0.841) |
0.717 | 0.804 | 0.594 | 0.736 (0.659–0.814) |
0.693 | 0.784 | 0.565 | 0.686 (0.606–0.767) |
0.629 | 0.745 | 0.464 | ||
| Intratumoral | 0.789 (0.721–0.858) |
0.708 | 0.796 | 0.586 | 0.755 (0.681–0.830) |
0.691 | 0.786 | 0.557 | 0.604 (0.491–0.718) |
0.554 | 0.368 | 0.815 | ||
| Intratumoral subregion | 0.750 (0.676–0.825) |
0.699 | 0.701 | 0.696 | 0.722 (0.644–0.800) |
0.675 | 0.691 | 0.652 | 0.568 (0.477–0.659) |
0.575 | 0.633 | 0.493 | ||
| Peritumoral | 0.754 (0.678–0.829) |
0.738 | 0.816 | 0.629 | 0.718 (0.638–0.798) |
0.720 | 0.796 | 0.614 | 0.604 (0.462–0.747) |
0.492 | 0.237 | 0.852 | ||
| Clinical | 0.779 (0.708–0.851) |
0.685 | 0.674 | 0.700 | 0.666 (0.583–0.750) |
0.643 | 0.643 | 0.643 | 0.549 (0.403–0.695) |
0.569 | 0.895 | 0.111 | ||
AUC, area under the curve; CI, confidence interval.
Among them, the intratumoral model was slightly better than the intratumoral subregion model in the training set, testing set, and external test set. The AUC of the intratumoral model, intratumoral subregion model, peritumoral model, and clinical data in the training set were 0.789 (95% CI: 0.721–0.858), 0.750 (95% CI: 0.676–0.825), 0.754 (95% CI: 0.678–0.829), and 0.779 (95% CI: 0.708–0.851) respectively. The AUC in the test set were 0.755 (95% CI: 0.681–0.830), 0.722 (95% CI: 0.644–0.800), 0.718 (95% CI: 0.638–0.798) and 0.666 (95% CI: 0.583–0.750) respectively. The AUC in the external test set was 0.604 (95% CI: 0.498–0.718), 0.568 (95% CI: 0.477–0.659), 0.604 (95% CI: 0.462–0.747), and 0.549 (95% CI: 0.403–0.695) respectively, indicating limited diagnostic efficacy. The performance of the intratumoral model, intratumoral subregion model and peritumoral model combined with clinical features slightly improved in both internal and external datasets. At the same time, the intratumoral model and intratumoral subregion model combined with the 2 mm peritumoral region showed enhanced diagnostic performance. The AUC of the intratumoral model combined with the 2 mm peritumoral region in the training set was 0.858 (95% CI: 0.808–0.907), the AUC in the test set was 0.789 (95% CI: 0.729–0.848), and the AUC in the internal test set was 0.808 (95% CI: 0.700–0.922). Meanwhile, we separately compared the performance of the intratumoral model versus the subregion model, the intratumoral + peritumoral model versus the subregion + peritumoral model, and the intratumoral + clinical model versus the subregion + clinical model. The results showed that across the training, testing, and test sets, the models incorporating intratumoral features consistently outperformed those utilizing subregional features. Based on these performances, we constructed a combination model of intratumoral, peritumoral, and clinical models, which showed the highest diagnostic performance. The model exhibited AUC values of 0.932, 0.833, and 0.741 in the training set, test set, and external test set, respectively, as shown in Figure 4. In the training set, the sensitivity, specificity, and accuracy were all greater than 0.85, in the test set, the sensitivity, specificity, and accuracy were all greater than 0.70, and in the external test set, the specificity was greater than 0.70.
Figure 4.
AUC values of the combined model (integrating intratumoral, peri-tumoral, and clinical models) across the internal training set, test set, and external test set. AUC, area under the curve; ROC, receiver operating characteristic.
In addition, the effectiveness of the combined models of intratumoral, peritumoral, and clinical models on training set, test set, and external test set datasets was rigorously evaluated through DCA and calibration plots. DCA demonstrated the clinical practicality of the model, especially at critical decision thresholds. For example, after the threshold exceeded 0.30, the training set, test set, and external test set all achieved the highest net benefit, indicating its outstanding potential in clinical applications. The calibration chart further supports the accuracy of the model. The predicted probability is highly consistent across all datasets, accurately reflecting the observed probability at different thresholds, further verifying its reliability and accuracy, as shown in Figure 5.
Figure 5.
Decision curve analysis and calibration curves for the best model (integrating intratumoral, peri-tumoral, and clinical models).
SHAP value analysis
We further elucidated the workings of the LR model through the application of SHAP analysis. This analysis significantly enhanced the interpretability of the machine learning model by providing clear insights into the influence that individual features exert on the predictive outcomes. The SHAP summary plot served as a powerful tool, Figure 6 shows the most influential features (gradient_firstorder_Skewness_Intra, square_glszm_SmallAreaLowGrayLevelEmphasis_Peri, lbp-3D k_glrIm_ShortRunEmphasis_Intra, age, wavelet-HHL_gldm_LargeDependenceLowGrayLevelEmphasis_Intra, were the top 5 features). It is worth noting that clinical features such as age, gender, and neurospecific_enolase exhibit a high level of influence. The detailed features were shown in Table S1.
Figure 6.
SHAP analysis was performed to further elucidate the working mechanism of the Logistic Regression model. SHAP, SHapley Additive exPlanations.
Discussion
Accurate differentiation between benign and malignant pulmonary nodules poses a significant challenge for clinicians, as diagnostic accuracy directly impacts the appropriateness of invasive procedure indications and the selection of optimal treatment timing (12).
In the baseline characteristic analysis, significant differences were observed between granulomatous lung disease and lung adenocarcinoma patients in terms of age, maximum tumor diameter, lobulation sign, gender and clarity of boundaries (P<0.05). The mean age of the lung adenocarcinoma group was significantly higher than that of the granulomatous lung disease group, and the maximum tumor diameter in the adenocarcinoma group was significantly larger compared to the granulomatous group. Additionally, lobulation sign was more prevalent in the adenocarcinoma group, consistent with prior studies (25-27). Compared with the granulomatous group, the adenocarcinoma group had a predominantly female composition and ill-defined tumor boundaries. The combination of age, sex, tumor size, lobulation sign, density uniform and clarity of tumor boundaries may optimize the benign-malignant prediction model. For instance, female patients over 60 years of age with nodules >15 mm in diameter, heterogeneous tumor density, ill-defined tumor boundaries, and lobulation sign exhibit an exceptionally high risk of malignancy, potentially warranting prioritized biopsy to guide specific clinical decisions.
This study sequentially utilized intratumoral radiomics features, peritumoral radiomics features, intratumoral subregional characteristics, and clinical risk factors to construct unimodal machine learning models for predicting granulomatous lung disease versus lung adenocarcinoma in solid pulmonary nodules. The unimodal models demonstrated suboptimal predictive performance, with AUC values of 0.789, 0.750, 0.754 and 0.779 in the training set, respectively, where the intratumoral model outperformed the intratumoral subregion model. To enhance model efficacy, we further combined features to construct new dual-modality models (intratumoral + clinical, intratumoral subregion + clinical, intratumoral + peritumoral, intratumoral subregion + peritumoral, peritumoral + clinical). This study found that all models showed improved performance after incorporating clinical or peritumoral features. These findings align with the results of Yang et al. (28) whose intratumoral-peritumoral combined model, developed using multicenter data, achieved AUC values of 0.906 and 0.886 in internal and external validations, respectively, confirming the advantage of integrating tumor region features in pulmonary nodule classification. Notably, the integration of peritumoral features with clinical factors AUC =0.897 (95% CI: 0.852–0.941) demonstrated a significant improvement compared to peritumoral features alone, indicating that clinical parameters can provide complementary biological context to imaging features. This supports the theory proposed by Chen et al. (10) that clinical variables hold critical supplementary value in enhancing the biological interpretation of radiomics characteristics.
Consequently, to maximize diagnostic capability, we developed a triple-feature model (intratumoral + peritumoral + clinical features), which achieved optimal performance in both the internal training and testing sets, with AUC values of 0.932 and 0.833, respectively. In the training set, sensitivity, specificity, and accuracy all exceeded 0.85. In the testing set, all three metrics exceeded 0.70; and in the external test set, specificity exceeded 0.70. This model significantly improved the discriminative power for distinguishing lung adenocarcinoma from granulomatous lung disease in solid pulmonary nodules.
Although the triple-feature model (intratumoral + peritumoral + clinical features) demonstrated excellent performance in the training set with AUC of 0.932, its efficacy diminished in external test with AUC of 0.741. In contrast, the dual-modality model (intratumoral + peritumoral features alone) exhibited more robust performance in external testing AUC =0.808. This discrepancy may stem from distribution shifts in clinical data within the external set. Despite the implementation of standardized preprocessing, inherent physical discrepancies between different scanner platforms, reconstruction algorithms, and original slice thicknesses may not be entirely eliminated. This objective factor partly explains why the more complex model integrating clinical features exhibited a more pronounced performance decline in external validation, whereas the imaging-only intratumoral-peritumoral dual-modality model demonstrated stronger cross-center stability (external test AUC =0.808). Under the realistically existing parameter variations, the current models—particularly the dual-modality imaging model—still display a certain degree of generalizability. However, their performance ceiling indicates that future multicenter studies need to further promote the standardization of prospective scanning protocols or develop more advanced domain adaptation algorithms. These findings suggest that, when clinical data quality is heterogeneous or uncontrolled, simplifying the model to a dual-modality imaging framework may enhance clinical practicality. This insight provides critical guidance for future multicenter study designs—for instance, prioritizing cross-center stability of imaging features first, followed by incremental integration of clinical data to refine predictive power.
This study applied habitat analysis to differentiate benign and malignant solid pulmonary nodules, overcoming the traditional “global averaging” approach for intratumoral feature extraction. However, the overall diagnostic performance of intratumoral subregional features remained suboptimal with AUC of 0.750, which may be attributed to two key factors. Firstly, limited imaging resolution hindered precise alignment with pathologically defined biological subregions, resulting in spatial mismatches between imaging and molecular profiles. This reduced the biological specificity of extracted features. Secondly, partitioning tumors into multiple subregions exponentially increased feature dimensionality without a proportional expansion in sample size, exacerbating data sparsity issues and elevating risks of model overfitting. Despite these limitations, the spatial heterogeneity distribution patterns of intratumoral subregional characteristics still demonstrate diagnostic value independent of conventional radiomics. This methodological innovation establishes a critical foundation for subsequent high-resolution multimodal studies to advance precision diagnostics (29,30).
There are several limitations in this study. Firstly, there are constraints in sample representativeness. The data were sourced from only two medical centers with a relatively small total sample size (n=230). Malignant nodules were exclusively limited to lung adenocarcinoma (58.3%), while benign nodules were restricted to inflammatory granulomas (41.7%). This incomplete coverage of pathological spectra (e.g., excluding squamous cell carcinoma, metastatic tumors and hamartomas) may limit the model’s generalizability to other lesion types. Future multicenter prospective studies should incorporate broader pathological diversity to establish a more comprehensive validation framework. Secondly, subregional model performance bottlenecks persist. Despite leveraging subregional imaging analysis to resolve spatial heterogeneity within tumors, the combined peritumoral-subregional model underperformed compared to the traditional intratumoral-peritumoral model (AUC =0.741 vs. 0.808). This may stem from insufficient sample size or the incomplete spatial differentiation of intratumoral heterogeneity in early-stage adenocarcinomas. Thirdly, the study relies heavily on ROI delineation. Manual ROI annotation is time-consuming and operator-dependent, potentially impacting radiomics feature reproducibility. To mitigate selection bias and enhance scalability, future work should integrate artificial intelligence (AI)-driven automated ROI segmentation tools, enabling standardized feature extraction for large-scale clinical deployment.
In summary, while habitat imaging did not outperform conventional intratumoral/peritumoral radiomics in this study, it represents a methodological advance toward capturing spatial heterogeneity. The suboptimal performance of subregion-based models highlights the practical challenges of implementing habitat analysis in routine CT-based nodule assessment. Future work should focus on improving spatial registration between imaging and pathology, integrating AI-driven automated subregion delineation, and validating habitat features in larger multi-center cohorts with diverse nodule types.
Conclusions
The multi-combination models developed in this study enhance the diagnostic accuracy for distinguishing benign from malignant solid pulmonary nodules, with the triple-feature model demonstrating the highest diagnostic performance. This approach has the potential to spare patients from unnecessary invasive procedures and strengthen clinical decision-making in the management of pulmonary nodules.
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 respective Ethics Committees of Shaoxing People’s Hospital (No. 2021-K-Y-021-01) and Shaoxing Second Hospital (No. 2024029001). Informed consent was waived in this retrospective study.
Footnotes
Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1869/rc
Funding: This study was supported by the Zhejiang Traditional Chinese Medicine Administration (No. 2025ZL139) and the Health and Wellness Technology Plan Project of Shaoxing City (No. 2024SKY092).
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1869/coif). This study was supported by the Zhejiang Traditional Chinese Medicine Administration (No. 2025ZL139) and the Health and Wellness Technology Plan Project of Shaoxing City (No. 2024SKY092). L.S. and Y.Y. are current employees of R&D Center, Hangzhou Deepwise. The authors have no other conflicts of interest to declare.
Data Sharing Statement
Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1869/dss
References
- 1.Siegel RL, Kratzer TB, Giaquinto AN, et al. Cancer statistics, 2025. CA Cancer J Clin 2025;75:10-45. 10.3322/caac.21871 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.de Koning HJ, van der Aalst CM, de Jong PA, et al. Reduced Lung-Cancer Mortality with Volume CT Screening in a Randomized Trial. N Engl J Med 2020;382:503-13. 10.1056/NEJMoa1911793 [DOI] [PubMed] [Google Scholar]
- 3.Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin 2024;74:12-49. 10.3322/caac.21820 [DOI] [PubMed] [Google Scholar]
- 4.Kratzer TB, Bandi P, Freedman ND, et al. Lung cancer statistics, 2023. Cancer 2024;130:1330-48. 10.1002/cncr.35128 [DOI] [PubMed] [Google Scholar]
- 5.Jeon H, Wang S, Song J, et al. Update 2025: Management of Non‑Small-Cell Lung Cancer. Lung 2025;203:53. 10.1007/s00408-025-00801-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Riely GJ, Wood DE, Ettinger DS, et al. Non-Small Cell Lung Cancer, Version 4.2024, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2024;22:249-74. 10.6004/jnccn.2204.0023 [DOI] [PubMed] [Google Scholar]
- 7.Jiang B, Han D, van der Aalst CM, et al. Lung cancer volume doubling time by computed tomography: A systematic review and meta-analysis. Eur J Cancer 2024;212:114339. 10.1016/j.ejca.2024.114339 [DOI] [PubMed] [Google Scholar]
- 8.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]
- 9.Aokage K, Suzuki K, Saji H, et al. Segmentectomy for ground-glass-dominant lung cancer with a tumour diameter of 3 cm or less including ground-glass opacity (JCOG1211): a multicentre, single-arm, confirmatory, phase 3 trial. Lancet Respir Med 2023;11:540-9. 10.1016/S2213-2600(23)00041-3 [DOI] [PubMed] [Google Scholar]
- 10.Chen C, Geng Q, Song G, et al. A comprehensive nomogram combining CT-based radiomics with clinical features for differentiation of benign and malignant lung subcentimeter solid nodules. Front Oncol 2023;13:1066360. 10.3389/fonc.2023.1066360 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Choi S, Yoon DW, Shin S, et al. Importance of Lymph Node Evaluation in ≤2-cm Pure-Solid Non-Small Cell Lung Cancer. Ann Thorac Surg 2024;117:586-93. 10.1016/j.athoracsur.2022.11.040 [DOI] [PubMed] [Google Scholar]
- 12.Peng M. Classification of pulmonary nodules in the era of precision medicine. Lancet Digit Health 2023;5:e633-4. 10.1016/S2589-7500(23)00154-1 [DOI] [PubMed] [Google Scholar]
- 13.Liu J, Qi L, Wang Y, et al. Development of a combined radiomics and CT feature-based model for differentiating malignant from benign subcentimeter solid pulmonary nodules. Eur Radiol Exp 2024;8:8. 10.1186/s41747-023-00400-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Mei M, Ye Z, Zha Y. An integrated convolutional neural network for classifying small pulmonary solid nodules. Front Neurosci 2023;17:1152222. 10.3389/fnins.2023.1152222 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Madariaga ML, Lennes IT, Best T, et al. Multidisciplinary selection of pulmonary nodules for surgical resection: Diagnostic results and long-term outcomes. J Thorac Cardiovasc Surg 2020;159:1558-1566.e3. 10.1016/j.jtcvs.2019.09.030 [DOI] [PubMed] [Google Scholar]
- 16.Williams BM, Herb J, Dawson L, et al. The Prevalence of Benign Pathology Following Major Pulmonary Resection for Suspected Malignancy. J Surg Res 2021;268:498-506. 10.1016/j.jss.2021.07.005 [DOI] [PubMed] [Google Scholar]
- 17.Chen M, Copley SJ, Viola P, et al. Radiomics and artificial intelligence for precision medicine in lung cancer treatment. Semin Cancer Biol 2023;93:97-113. 10.1016/j.semcancer.2023.05.004 [DOI] [PubMed] [Google Scholar]
- 18.Warkentin MT, Al-Sawaihey H, Lam S, et al. Radiomics analysis to predict pulmonary nodule malignancy using machine learning approaches. Thorax 2024;79:307-15. 10.1136/thorax-2023-220226 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Xiao Y, Yu D. Tumor microenvironment as a therapeutic target in cancer. Pharmacol Ther 2021;221:107753. 10.1016/j.pharmthera.2020.107753 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Elhanani O, Ben-Uri R, Keren L. Spatial profiling technologies illuminate the tumor microenvironment. Cancer Cell 2023;41:404-20. 10.1016/j.ccell.2023.01.010 [DOI] [PubMed] [Google Scholar]
- 21.Prior O, Macarro C, Navarro V, et al. Identification of Precise 3D CT Radiomics for Habitat Computation by Machine Learning in Cancer. Radiol Artif Intell 2024;6:e230118. 10.1148/ryai.230118 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Wu F, Fan J, He Y, et al. Single-cell profiling of tumor heterogeneity and the microenvironment in advanced non-small cell lung cancer. Nat Commun 2021;12:2540. 10.1038/s41467-021-22801-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Shi Z, Huang X, Cheng Z, et al. MRI-based Quantification of Intratumoral Heterogeneity for Predicting Treatment Response to Neoadjuvant Chemotherapy in Breast Cancer. Radiology 2023;308:e222830. 10.1148/radiol.222830 [DOI] [PubMed] [Google Scholar]
- 24.Chen H, Liu Y, Zhao J, et al. Quantification of intratumoral heterogeneity using habitat-based MRI radiomics to identify HER2-positive, -low and -zero breast cancers: a multicenter study. Breast Cancer Res 2024;26:160. 10.1186/s13058-024-01921-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Han D, Zhao J, Hao S, et al. Integrative radiomics analysis of peri-tumoral and habitat zones for predicting major pathological response to neoadjuvant immunotherapy and chemotherapy in non-small cell lung cancer. Transl Lung Cancer Res 2025;14:1168-84. 10.21037/tlcr-2024-1131 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Duan L, Liu W, Li M, et al. CT radiomics from intratumor and peritumor regions for predicting poorly differentiated invasive nonmucinous pulmonary adenocarcinoma. Sci Rep 2025;15:14434. 10.1038/s41598-025-99465-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Yang L, Jiang Z, Tong J, et al. Development and validation of a preoperative CT‑based radiomics nomogram to differentiate tuberculosis granulomas from lung adenocarcinomas: an external validation study. BMC Cancer 2024;24:670. 10.1186/s12885-024-12422-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Yang Z, Dong H, Fu C, et al. A nomogram based on CT intratumoral and peritumoral radiomics features preoperatively predicts poorly differentiated invasive pulmonary adenocarcinoma manifesting as subsolid or solid lesions: a double-center study. Front Oncol 2024;14:1289555. 10.3389/fonc.2024.1289555 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Jirapatnakul A, Yip R, Myers KJ, et al. Assessing the impact of nodule features and software algorithm on pulmonary nodule measurement uncertainty for nodules sized 20 mm or less. Quant Imaging Med Surg 2024;14:5057-71. 10.21037/qims-23-1501 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Yu S, Yang Y, Wang Z, et al. CT-based conventional radiomics and quantification of intratumoral heterogeneity for predicting benign and malignant renal lesions. Cancer Imaging 2024;24:130. 10.1186/s40644-024-00775-8 [DOI] [PMC free article] [PubMed] [Google Scholar]