Improving Risk Stratification of Pulmonary Nodules: An Integrated Perinodular Vascular and Radiomic Model for Clinical Decision Support

Improving Risk Stratification of Pulmonary Nodules: An Integrated Perinodular Vascular and Radiomic Model for Clinical Decision Support | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Improving Risk Stratification of Pulmonary Nodules: An Integrated Perinodular Vascular and Radiomic Model for Clinical Decision Support Wen Qiu, Chunyi Liang, Wenxuan Luo, Jinxiu Lin, Jianli Cao, Peng Peng, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8175303/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 10 Apr, 2026 Read the published version in BMC Medical Imaging → Version 1 posted 12 You are reading this latest preprint version Abstract Objectives Lung cancer is the leading cause of cancer-related deaths worldwide, with most patients diagnosed at advanced stages. Accurate differentiation of benign and malignant pulmonary nodules remains a major clinical challenge. Materials and Methods We established and validated the perinodular vessel count (PVC) as an instrumental imaging biomarker, demonstrating its significant contribution to discriminating malignant pulmonary nodules. Leveraging this finding, we constructed an integrated predictive model incorporating intranodular and perinodular radiomics, PVC, and relevant clinical variables. A two-tiered feature selection strategy employing both MRMR and Relief algorithms was implemented to refine feature sets, followed by the development of an ensemble decision tree-based classifier. The model underwent rigorous multi-center validation and exhibited diagnostic performance on par with that of three experienced clinicians, underscoring its potential utility in clinical decision-making. Results The model incorporating perinodular vascular features significantly outperformed the non-vascular feature model. Furthermore, the combined clinical-vascular-radiomic model demonstrated substantially improved performance over the clinical-vascular model, achieving AUCs of 0.8704 (CI: [0.6417,0.9676]) validation set, 0.8225 on independent test set (CI: [0.7298,0.9168]), and 0.7937 (CI: [0.4234,1]) on external test set. The PVC feature was consistently identified as one of the most important feature among all features in both feature selection and SHAP interpretability analysis. Conclusion Integration of vascular characteristics markedly improves diagnostic performance and model generalizability. The consistent importance of PVC highlights its clinical value, and the model shows promising potential to assist in decision-making and reduce unnecessary invasive procedures. Computed tomography Benign-Malignant Classification Pulmonary Nodules Perinodular Vascular Count Combined Clinical-Radiomic-Features Model Feature Interpretability Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Key points Question: To achieve the early and accurate classification of pulmonary nodules, developing models based on more robust perinodular nodule features is crucial. Findings: A combined model utilizing clinical and radiomic features, particularly perinodular vascular count, enhances the accuracy and generalizability of pulmonary nodule evaluation on multicenter dataset. Clinical Relevance: This study offers a clinical decision-support tool for clinicians to reduce overtreatment by reliably ruling out malignancy in ambiguous pulmonary nodules, thus preventing unnecessary invasive procedures and improving patient care. Introduction Lung cancer has reemerged as the most commonly diagnosed cancer worldwide, with 2.481 million new cases in 2022 accounting for 12.4% of all cancer diagnoses [ 1 , 2 ]. Moreover, it is projected to become the leading cause of global cancer incidence and mortality by 2050 [ 3 ]. Prognosis is highly stage-dependent: the five-year survival rate is 77–92% for Stage I but plummets to 0–36% for Stage II, underscoring the critical importance of early detection [ 4 ]. However, a majority of lung cancer patients are diagnosed at advanced stages, missing the window for curative treatment. A pulmonary nodule (PN) is a focal, round, high-density shadow in the lung, either solid or subsolid, with a diameter of ≤ 30 mm. It serves as a "carrier" or early manifestation of lung cancer, with the vast majority of lung cancers first detected as nodules. Nevertheless, most pulmonary nodules are not cancerous but are caused by benign conditions such as infection or inflammation. Consequently, the clinical management strategy for a pulmonary nodule depends directly on whether it is benign or malignant. The International Agency for Research on Cancer provides clear pathological criteria. For nodules pathologically identified as benign, such as preinvasive glandular lesions, active surveillance is typically recommended. In contrast, malignant nodules generally require surgical resection [ 5 , 6 ]. To reduce over-treatment rates without compromising the malignant detection rate, thus allowing for effective stratified management of nodules—has become a major clinical challenge. The imprecision of conventional radiological assessment in characterizing pulmonary nodules can result in both delayed treatment and unnecessary procedures. [ 7 ] Radiomics and other AI approaches have shown promise in improving diagnostic accuracy by extracting high-dimensional quantitative features from medical images. [ 8 – 10 ] These intranodular features can capture tumor heterogeneity and have proven diagnostic value. However, tumors do not exist in isolation. Evidence shows that perinodular characteristics, like changes in surrounding vasculature, are key markers of tumor aggressiveness and prognosis, highlighting a vital direction for AI-driven diagnosis. [ 11 , 12 ] Previous research on lung nodule diagnosis has largely examined intranodular radiomics, perinodular features, and clinical factors in isolation, with a notable scarcity of integrated models and a lack of quantitative validation for perinodular characteristics. To address this gap, our study proposes a comprehensive model that integrates these three domains to enhance diagnostic precision. Moreover, a key objective is the interpretable quantification of PVC. We anticipate that this holistic strategy will yield deeper insights into nodule behavior and directly support improved clinical decision-making. Materials and Methods This study aims to develop a machine learning model integrating multimodal features, which will subsequently be evaluated on multi-center test sets to assess its generalizability. Furthermore, a rigorous interpretability analysis was conducted to validate the importance of the PVC feature. The specific workflow is outlined in Fig. 1 . Dataset This retrospective study curated a primary dataset of 218 nodules from 100 patients using CT scans acquired between January 2023 and December 2024 from hospital 1. An independent internal test set of 81 nodules from 62 patients was formed from scans dated January to July 2025, and an external test set of 21 nodules from 16 patients was included from hospital 2. All cases met the following criteria: (a) availability of postoperative pathological results; (b) nodule maximum diameter ≤ 30 mm; (c) CT scan resolution ≤ 1 mm and slice thickness < 1.5 mm; and (d) availability of essential clinical characteristics. The detailed distribution of the data is shown in Table 1 . Table 1 presents the distribution of important features related to the data in the training set, validation set, and two test sets. Variable Training Validation Independent Test External Test No. of Lesions 186 32 81 21 Age 57.01 ± 12.43 58.06 ± 12.81 60.22 ± 11.73 59.33 ± 11.99 Gender Male 122 16 32 12 Female 64 16 49 9 PVC 2.59 ± 1.59 2.94 ± 1.72 2.65 ± 1.39 3.1 ± 2 Homogeneous density 66% 66% 47% 48% Nodule type Solid 23 1 18 9 Part-Solid 54 12 25 4 Ground-Glass 109 19 36 8 Radiological signs Lobulation 23% 25% 31% 48% Pleural indentation 22% 38% 38.3% 67% Vascular convergence 7% 16% 22.2% 43% Vacuole 24% 16% 20% 24% Feature Extraction and Selection Three categories of features were extracted from each pulmonary nodule for comprehensive analysis. All features included are routinely utilized as standard measures in clinical and research settings. The first category comprised 30 clinical, laboratory, and radiological features, including patient demographics (age, gender, cough, sputum, hemoptysis, fever, chest pain), medical history (smoking, diabetes, tuberculosis, cancer, family), white blood cell count, carcinoembryonic antigen level, nodule morphological characteristics(nodule type, long-axis diameter, long-axis to short-axis ratio, volume, homogeneous density, boundary clarity, morphological regularity), imaging signs (lobulation, spiculation, pleural indentation, vascular convergence, vacuole sign, calcification, cavity), and key imaging signs—with perinodular vessel count (PVC) being a focus of this study. [ 13 – 15 ] For the PVC feature, we established a criterion in Fig. 2 and performed the annotation based on this standard. Clinical and laboratory features were obtained through retrospective review of patient outpatient and laboratory records. Each imaging sign for every nodule was independently annotated by two resident physicians, with any discrepancies adjudicated by a senior attending radiologist. Figure 2. Identification of perinodular vessel counts: (A) Vessels surrounding the tumor: vessel (a, d, e) contact with the tumor on the hilar side and are identified as proximal marginal vessels, vessels (b, c) connect to the tumor only on the pleural side are thus classified as distal marginal vessels. Only proximal marginal vessels are counted to PVC. (B) Schematic representation of a subpleural tumor, in which only proximal vessels are visible. (C–F) Axial (C), coronal (D), and volumetric reconstruction images (E, F) from the same patient depict the proximal vessels adjacent to the tumor. Three proximal marginal vessels are shown in contact with the tumor (red arrows). The second and third categories consisted of 1,927 radiomic features each, extracted from the core nodule region and a 3-voxel dilated perinodular region, respectively, where each region of interest (nodule) was meticulously delineated. [ 16 , 17 ] These radiomic features encompassed intensity, texture, and shape information, and were automatically generated using the United Imaging Research Platform following semi-automatic segmentation. To address high dimensionality and potential redundancy in the initial feature set (~ 4,000 features), a two-stage filter-based selection strategy was implemented. The Maximum Relevance Minimum Redundancy (MRMR) [ 18 ] algorithm was first applied for global redundancy reduction, prioritizing features with high relevance to malignancy and low inter-feature redundancy based on mutual information. This was followed by the Relief algorithm [ 19 ] to refine local discriminatory power by evaluating feature effectiveness in distinguishing benign and malignant cases within local instance neighborhoods. This combined approach effectively balanced redundancy control with classification boundary optimization while preserving predictive performance and interpretability. In subsequent modeling, feature selection was adapted to the input type. For clinical-only models, Relief was used to select the top 30% of features (approximately 9 features). For combined clinical-radiomic models, MRMR first reduced the feature set to the top 20%, from which Relief then selected a final set of 18 features—double the number used in clinical-only models—to fully leverage multimodal information. Classification This study was designed with a systematic experimental protocol to comprehensively evaluate the performance, robustness, and practical potential of the proposed multimodal classification model in real-world clinical settings. The core objectives included: (1) validating the necessity and contribution of the key clinical feature PVC to model performance; (2) elucidating the synergistic effects and individual value of features derived from different sources (clinical features, intranodular radiomic features, and perinodular radiomic features); and (3) assessing the model’s advantage over pure human diagnosis and its potential to reduce the rate of overtreatment. All experiments were conducted using rigorously partitioned training, validation, and independent test sets. First, to accurately evaluate the role of perinodular vessel count in the classification task, an ablation study at the feature level was designed. By comparing models trained on the full feature set and model intentionally ablated by excluding the vessel count feature, the necessity and independent contribution of this feature to overall performance were quantified. Second, to dissect the synergistic mechanisms among features from different sources, a progressive feature integration strategy was adopted. A baseline model relying solely on structured clinical features was initially established. Intranodular radiomic features extracted from the target region of interest (ROI) were then incrementally incorporated to evaluate their added value beyond clinical information, which is the Clinic_Intranodular_Radiomics (CIR) model. Finally, perinodular radiomic features were encoded and integrated to construct a comprehensive multimodal fusion model, which is the Clinic_Intranodular_Perinodular_Radiomics (CIPR) model, thoroughly validating the complementarity and integrative efficacy of cross-source features. Lastly, a controlled experiment involving specialized physicians was conducted to quantify the auxiliary value of the model in real clinical practice. The physician group performed independent diagnoses without assistance and recorded their conclusions and the model group generated predictions and confidence scores autonomously. Comparisons among these two groups were made along two dimensions—diagnostic sensitivity and specificity—to elucidate the practical advantages of human-AI collaborative decision-making. The primary dataset was randomly partitioned into a training set and an internal test set in an 85:15 ratio. Both the independent and external test sets were strictly designed to include patients and nodules distinct from those in the training set, ensuring robust evaluation of model generalizability. For model training, all experiments were conducted using a bagging ensemble of decision trees [ 20 , 21 ] implemented on the uAI Research Portal (uAI Research Portal, Version: 20240130) [ 22 ]. Specifically, the Scikit-Learn library, an open-source machine learning library, was utilized to build the model. The number of trees was set to half the number of features to avoid overfitting, and the tree depth was set to the platform’s default value. Feature Interpretability Analysis A multi-level analysis was employed to interpret the selected features. First, PVC’s importance was objectively assessed using MRMR (reflecting feature relevance and redundancy) and Relief (confirming classification utility) algorithms. Second, SHAP analysis globally quantified the contribution of PVC to model predictions, illustrating how its variation affects output probabilities. [ 23 , 24 ] Statistical Analysis Model predictions were evaluated using confusion matrix-derived performance metrics, including Accuracy, Sensitivity, Specificity, Youden Index, and the F1-Score, to comprehensively assess classification efficacy. Given the limited sample size, bootstrap resampling [ 25 , 26 ] (100 iterations) was applied on results of each test set to estimate confidence intervals for each metric. Inter-group comparisons were performed using Student’s t-test [ 27 ], with a significance threshold set at α = 0.05. Key results are reported as mean and confidence interval. All analyses were conducted on both validation and test sets, including the internal validation set, the internal independent test set, and the external test set, to ensure unbiased evaluation. Results Classification Results As illustrated in Fig. 3 (a) and Table 2 , an ablation experiment focusing on the perinodular vessel count (PVC) demonstrated that incorporating this feature into the feature set—followed by feature selection—significantly enhanced the AUC to 0.8178, markedly outperforming the model trained without PVC. Correspondingly, sensitivity, specificity, and F1-score improved substantially from 0.6316, 0.6154, and 0.6667 to 0.8421, 0.7692, and 0.8421, respectively. These results indicate that the inclusion of PVC significantly boosts the diagnostic performance of conventional clinical features in distinguishing benign from malignant lesions. Table 2 Model Comparison of the Malignant-Benign Classification Ability of PVC Features AUC Sen Spe Acc Pre F1 Clinic 0.8178 (0.6417, 0.9676) 0.8421 (0.6581, 1.0000) 0.7692 (0.5158, 1.0000) 0.8125 (0.6562, 0.9375) 0.6113 (0.3133, 0.8693) 0.8421 (0.6667, 0.9511) CIPR 0.7389 *** (0.5448, 0.9076) 0.6316 *** (0.4148, 0.8534) 0.6154 *** (0.3343, 0.8182) 0.6250 *** (0.4523, 0.7812) 0.2470 *** (-0.0873, 0.5918) 0.6667 *** (0.4823, 0.8392) Note : Significance levels are denoted as follows: *p < 0.05, **p < 0.01, ***p < 0.001 In the multi-source feature ablation experiments, as shown in Fig. 3 (b) and Table 3 , the integration of radiomic features led to a notable increase in AUC from 0.8178 to 0.8704. Furthermore, Table 3 and Fig. 3 reveal that the simultaneous incorporation of both intranodular and perinodular radiomic features yielded significant improvements in AUC and specificity compared to using intranodular features alone. We also compared clinic model against CIPR model on two independent test sets, demonstrated in Fig. 4 and Table 4 . The CIPR model consistently outperformed the clinic model across both datasets. Notably, on the Yangjiang independent test set, the CIPR model achieved significantly higher values across all classification metrics. Specifically, it exhibited superior sensitivity and F1-scores on both datasets, while matching or exceeding the clinic model in other performance indicators. These findings underscore the substantial enhancement in discriminative efficacy brought by radiomic features, particularly those related to perinodular vascular patterns. Table 3 Ablation study of clinical features including PVC, intranodular and perinodular radiomics AUC Sen Spe Acc Youden F1 CIPR 0.8704 (0.6417, 0.9676) 0.8421 (0.6581, 1.0000) 0.7692 (0.5158, 1.0000) 0.8125 (0.6562, 0.9375) 0.6113 (0.3133, 0.8693) 0.8421 (0.6667, 0.9511) CIR 0.8158 *** (0.6397, 0.9543) 0.8947 (0.7099, 1.0000) 0.6923 *** (0.5158, 1.0000) 0.8125 (0.6398, 0.9375) 0.5870 (0.2297, 0.8745) 0.8500 (0.6766, 0.9545) Clinic 0.8178 *** (0.6417, 0.9676) 0.8421 (0.6581, 1.0000) 0.7692 (0.5158, 1.0000) 0.8125 (0.6562, 0.9375) 0.6113 (0.3133, 0.8693) 0.8421 (0.6667, 0.9511) Note : Significance levels are denoted as follows: *p < 0.05, **p < 0.01, ***p < 0.001 Table 4 Comparison of classification results between CIPR and clinic model on independent and external test dataset AUC Sen Spe Acc Youden F1 Independent Testing Dataset CIPR 0.8225 (0.7298, 0.9168) 0.8824 (0.8059, 0.9622) 0.7000 (0.5601, 0.8571) 0.8148 (0.7343, 0.8889) 0.5824 (0.4088, 0.7563) 0.8571 (0.7770, 0.9167) Clinic 0.8007 ** (0.6946, 0.9025) 0.8235 *** (0.7315, 0.9262) 0.6667 ** (0.4835, 0.8458) 0.7654 *** (0.6725, 0.8519) 0.4902 *** (0.2728, 0.6929) 0.8155 *** (0.7368, 0.8879) External Dataset CIPR 0.7937 (0.4234, 1.0000) 0.8750 (0.6750, 1.0000) 0.8 (0.3786, 1.0000) 0.8571 (0.6667, 1.0000) 0.6750 (0.2181, 1.0000) 0.9032 (0.7586, 1.0000) Clinic 0.7562 (0.3438, 1.0000) 0.8125 *** (0.5714, 1.0000) 0.8 (0.3786, 1.0000) 0.8095 *** (0.6190, 0.9524) 0.6125 * (0.1389, 0.9460) 0.8667 *** (0.6804, 0.9722) Note : Significance levels are denoted as follows: *p < 0.05, **p < 0.01, ***p < 0.001 Figure 4 further demonstrates that the model overall matched or exceeded the diagnostic performance of two senior radiologists and one junior radiologist on the independent internal test set. Although two radiologists achieved specificities above 0.8, their sensitivities were both below 0.7, considerably lower than the model’s sensitivity of 0.8824. Moreover, in terms of the more balanced metric—Youden Index—the model attained a value of 0.5824 on the internal test set, outperforming the three radiologists’ scores of 0.4667, 0.5627, and 0.4608, respectively. On the external test set, the model achieved a Youden Index of 0.6750, significantly surpassing the radiologists’ results of 0.4875, 0.0125, and 0.225. These results indicate that, compared with human evaluators, the model exhibits superior balance between sensitivity and specificity, thereby assisting clinicians in avoiding overdiagnosis while maintaining a high malignant detection rate. Interpretability Analysis of PVC Our analysis reaffirms that the PVC feature, along with other medical imaging-related characteristics, plays a pivotal role in significantly enhancing the diagnostic performance of the model. Subsequent interpretability analyses—including feature importance metrics and SHAP analysis—consistently underscore the critical importance of the PVC feature. First, examination of feature distributions within the training set reveals a statistically significant disparity in PVC values between benign and malignant cases, as illustrated in Fig. 5 (a). During feature selection, the MRMR algorithm assigned the highest score to the PVC feature, indicating that it achieves an optimal balance between maximal relevance to the target variable and minimal redundancy with other features (Fig. 5 (b)). In other words, the algorithm identifies PVC as providing the greatest information gain with the least overlap. In a subsequent feature ranking using the Relief algorithm, PVC was considered as the second-highest importance score, immediately following “Homogeneous Density”, further corroborating its discriminative power in classifying benign and malignant nodules (Fig. 5 (c)). SHAP analysis likewise highlighted the prominence of the PVC feature. As shown in Fig. 5 (d), PVC attained the second-highest mean |SHAP value|, confirming its role as one of the most influential predictors at both local and global interpretation levels. This offers robust and explainable evidence that the model heavily relies on this feature during decision-making. Furthermore, the SHAP dependence plot [ 28 ] in Fig. 5 (e) demonstrates that higher PVC values correlate with increased model tendency toward predicting malignancy—a finding consistent with clinical intuition. Additionally, Homogeneous Density also emerged as a major contributor. The SHAP interaction plot [ 29 ] in Fig. 5 (f) reveals notable interaction effects between these two features: particularly in cases with non-uniform density, PVC serves as an even more decisive factor. In contrast, when nodule density is uniform, their joint influence approaches zero, implying that additional features are required to guide the model’s decision. In conclusion, the PVC feature constitutes a cornerstone of the model’s ability to differentiate between benign and malignant nodules. As an interpretable decision-support tool, it holds considerable promise for integration into clinical workflows, thereby aiding diagnosticians in making more informed and reliable decisions. Discussion Early and accurate classification of pulmonary nodules based on conventional CT features is clinically challenging due to nonspecific imaging appearances, technical variability, and limited observer agreement. Current criteria—such as size, density, and morphology—often lead to diagnostic overlap and high false-positive rates, particularly in intermediate-sized nodules [ 30 ]. Furthermore, acquisition parameters (e.g., slice thickness, dose) significantly affect diagnostic outputs [ 31 ], and inter-observer consistency remains moderate (κ = 0.454) [ 32 ]. Similarly, AI models continue to produce false results and exhibit bias, restricting clinical utility [ 33 , 34 ]. This study therefore seeks to identify more robust discriminative features to improve nodule assessment. This study confirms the diagnostic significance of PVC and establishes that a combined model (CIPR) incorporating clinical and radiomic features—particularly those capturing perinodular vascular geometry—achieves more accurate and generalizable performance than clinical models alone. Furthermore, the AI approach provided more consistent and objective assessments compared to physician judgments, highlighting its potential as a decision-support tool to mitigate overtreatment in cases of diagnostic uncertainty. Clinically, our model shows strong potential to reduce over-treatment. According to all datasets, nodules with ≥ 4 perinodular vessels were rarely benign (4/78), indicating low over-treatment risk. More importantly, in cases with fewer vessels—where visual assessment is difficult—the model provided high specificity (> 70%) without compromising sensitivity (≥ 70%), as shown in Table 5 , enabling more confident rule-out of malignancy and avoiding unnecessary procedures in ambiguous cases. Table 5 CIPR model classification results on nodules with PVC < 4 Dataset Amount of PVC < 4 Sensitivity Specificity Validation 23 0.7272 0.8333 Independent Test 59 0.7931 0.7 External Test 16 0.7272 0.8 This study has several limitations. The sample size, particularly the small number of benign cases (n = 5) in the external test set, may affect the robustness of the evaluation. Furthermore, interpreter subjectivity affects vessel count assessment due to differences in physician qualifications and image quality, potentially influencing model predictions. Future work should focus on establishing a unified vessel counting protocol, elucidating the pathological basis of perinodular vascular features, and developing automated quantification methods to improve objectivity and clinical integration. Declarations Acknowledgements Thanks to Central Research Institute, United Imaging Healthcare for their technical support. We also thank Guangdong Medical University Affiliated Hospital for providing the external test dataset.We also extend our heartfelt appreciation to the anonymous reviewers, whose diligent efforts significantly enhanced the quality of this paper. Author contributions W Q conceptualized the study, curated the data, and conducted formal analysis along. W Q, C-Y L, and W-X L were responsible for the investigation. The methodology was developed by W Q and Y-F H. Project administration was managed by W Q. Resources were provided by W Q, W-X L, J-X L, J-L C, P Pand Y-H Z. Software development was handled by Y-F H. Supervision and validation were conducted by Y-F H and R-Z C. Visualization was carried out by W Q. W Q and Y-F H prepared the original draft, while Y-F H and R-Z C reviewed and edited the manuscript. All authors read and approved the final manuscript. Funding This study has received funding by the Yangjiang Science and Technology Bureau,China(Grant No. SF2025055). Ethics approval and consent to participate This study was approved by the Ethics Committee of Yangjiang People’s Hospital (Approval No.KY-2025-42-1). The Committee waived the requirement for patient consent for this investigation. The authors declare that the research involving human data complies with the Helsinki Declaration. Consent to Publish declaration Not applicable. Data Availability declaration All data and tables included in this study are available from the corresponding author upon reasonable request. Conflicts of Interest The authors declare that they have no conflicts of interest. References Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74(3):229–63. 10.3322/caac.21834 . Wang C, Shao J, Song L, Ren P, Liu D, Li W. Persistent increase and improved survival of stage I lung cancer based on a large-scale real-world sample of 26,226 cases. Chin Med J (Engl). 2023;136(16):1937–48. 10.1097/CM9.0000000000002729 . Bizuayehu HM, Ahmed KY, Kibret GD, et al. Global Disparities of Cancer and Its Projected Burden in 2050. JAMA Netw Open. 2024;7(11):e2443198. 10.1001/jamanetworkopen.2024.43198 . Published 2024 Nov 4. Chinese Thoracic Society. Chinese expert consensus on diagnosis of early lung cancer (2023 Edition). Zhonghua Jie He He Hu Xi Za Zhi. 2023; 46:1–18. 10.3760/cma.j.cn112147-20220712-00592 Tsao MS, Nicholson AG, Maleszewski JJ, Marx A, Travis WD. Introduction to 2021 WHO Classification of Thoracic Tumors. J Thorac Oncol. 2022;17(1):e1–4. 10.1016/j.jtho.2021.09.017 . Wang W, Yang B, Wu H, et al. Auxiliary Diagnosis of Pulmonary Nodules' Benignancy and Malignancy Based on Machine Learning: A Retrospective Study. J Multidiscip Healthc. 2025;18:3735–48. 10.2147/JMDH.S518166 . Published 2025 Jun 27. Chen D, Yang L, Zhang W, Shen J, Van Schil PEY, Divisi D, Seetharamu N, Gu J. Prevalence and management of pulmonary nodules: a systematic review and meta-analysis. J Thorac Dis. 2024;16(7):4619–32. 10.21037/jtd-24-874 . Epub 2024 Jul 24. PMID: 39144359; PMCID: PMC11320231. Ayasa Y, Alajrami D, Idkedek M, Tahayneh K, Akar FA. The Impact of Artificial Intelligence on Lung Cancer Diagnosis and Personalized Treatment. Int J Mol Sci. 2025;26(17):8472. 10.3390/ijms26178472 . Published 2025 Aug 31. Wang C, Shao J, He Y, et al. Data-driven risk stratification and precision management of pulmonary nodules detected on chest computed tomography. Nat Med. 2024;30:3184–95. https://doi.org/10.1038/s41591-024-03211-3 . Yang Y, Wang P, Yu C, et al. Application of artificial intelligence medical imaging aided diagnosis system in the diagnosis of pulmonary nodules. BMC Med Inf Decis Mak. 2025;25(185). https://doi.org/10.1186/s12911-025-03009-4 . Zhang XC, Lv FJ, Fu BJ, Liang ZR, Chu ZG. Significance of marginal vessels in differentiating peripheral small-cell lung cancer and benign lung tumor. Acta Radiol. 2023;64(9):2526–34. 10.1177/02841851231188060 . Wang W, Qi X, He Y, et al. Radiomics of Vascular Structures in Pulmonary Ground-Glass Nodules: A Predictor of Invasiveness: Radiomics of Vascular Structures in GGNs for Tumor Invasiveness Prediction. Curr Med Imaging. 2025. 10.2174/0115734056385352250410053810 . MacMahon H, Naidich DP, Goo JM, et al. Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017. Radiology. 2017;284(1):228–43. 10.1148/radiol.2017161659 . Baeza S, Gil D, Sanchez C, et al. Radiomics and Clinical Data for the Diagnosis of Incidental Pulmonary Nodules and Lung Cancer Screening: Radiolung Integrative Predictive Model. Arch Bronconeumol. 2024;60(Suppl 2):S22–30. 10.1016/j.arbres.2024.05.027 . Zwirewich CV, Vedal S, Miller RR, Müller NL. Solitary pulmonary nodule: high-resolution CT and radiologic-pathologic correlation. Radiology. 1991;179(2):469–76. 10.1148/radiology.179.2.2014294 . Selvam M, Chandrasekharan A, Sadanandan A, et al. Radiomics as a non-invasive adjunct to Chest CT in distinguishing benign and malignant lung nodules. Sci Rep. 2023;13(19062). https://doi.org/10.1038/s41598-023-46391-7 . Masquelin AH, Alshaabi T, Cheney N, Estépar RSJ, Bates JHT, Kinsey CM. Perinodular Parenchymal Features Improve Indeterminate Lung Nodule Classification. Acad Radiol. 2023;30(6):1073–80. 10.1016/j.acra.2022.07.001 . Peng H, Long F, Ding C. Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell. 2005;27(8):1226–38. Kim K, Rendell LA. Morgan Kaufmann, Los Altos, CA,. A practical approach to feature selection, in: Proceedings 9th international Conference on Machine Learning, Aberdeen, Scotland (1992). Altman N, Krzywinski M. Ensemble methods: bagging and random forests. Nat Methods. 2017;14:933–4. https://doi.org/10.1038/nmeth.4438 . Plaia A, Buscemi S, Fürnkranz J, et al. Comparing Boosting and Bagging for Decision Trees of Rankings. J Classif. 2022;39:78–99. https://doi.org/10.1007/s00357-021-09397-2 . Wu J, Xia Y, Wang X, et al. uRP: An integrated research platform for one-stop analysis of medical images. Front Radiol. 2023;3:1153784. 10.3389/fradi.2023.1153784 . Published 2023 Apr 18. Lundberg S, Lee S. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17). 2017; Curran Associates Inc., Red Hook, NY, USA, 4768–4777. Ponce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024;17(11):e70056. 10.1111/cts.70056 . PMID: 39463176; PMCID: PMC11513550. Efron B. Nonparametric estimates of standard error: The jackknife, the bootstrap and other methods. Biometrika. 1981;68:589–99. Cho K. Performance assessment through bootstrap. IEEE Trans Pattern Anal Mach Intell. 1997;19:1185–98. Fisher RA. Applications of ‘Student’s’ distribution. Metron. 1925;5:90–104. Hajjar M, Aldabbagh G, Albaradei S. Interpretable Multi-Cancer Early Detection Using SHAP-Based Machine Learning on Tumor-Educated Platelet RNA. Diagnostics (Basel). 2025;15(17):2216. Published 2025 Sep 1. 10.3390/diagnostics15172216 Fumagalli F, Muschalik M, Kolpaczki P, Hüllermeier E, Hammer BE. SHAP-IQ: Unified approximation of any-order shapley interactions. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023), 2023. Yao Y, Wang X, Guan J et al. Metabolomic differentiation of benign vs malignant pulmonary nodules with high specificity via high-resolution mass spectrometry analysis of patient sera. Nat Commun. 2023;14(1):2339. Published 2023 Apr 24. 10.1038/s41467-023-37875-1 Wang C, Shao J, He Y, et al. Data-driven risk stratification and precision management of pulmonary nodules detected on chest computed tomography. Nat Med. 2024;30(11):3184–95. 10.1038/s41591-024-03211-3 . Martini K, Ottilinger T, Serrallach B, et al. Lung cancer screening with submillisievert chest CT: Potential pitfalls of pulmonary findings in different readers with various experience levels. Eur J Radiol. 2019;121:108720. 10.1016/j.ejrad.2019.108720 . Zhao M, Xue G, He B, et al. Integrated multiomics signatures to optimize the accurate diagnosis of lung cancer. Nat Commun. 2025;16(1):84. 10.1038/s41467-024-55594-z . Sourlos N, Wang J, Nagaraj Y, van Ooijen P, Vliegenthart R. Possible Bias in Supervised Deep Learning Algorithms for CT Lung Nodule Detection and Classification. Cancers (Basel). 2022;14(16):3867. 10.3390/cancers14163867 . Published 2022 Aug 10. Additional Declarations No competing interests reported. Supplementary Files SupplementaryFiles.docx Cite Share Download PDF Status: Published Journal Publication published 10 Apr, 2026 Read the published version in BMC Medical Imaging → Version 1 posted Editorial decision: Revision requested 16 Feb, 2026 Reviews received at journal 13 Feb, 2026 Reviews received at journal 07 Feb, 2026 Reviews received at journal 06 Feb, 2026 Reviewers agreed at journal 06 Feb, 2026 Reviewers agreed at journal 04 Feb, 2026 Reviewers agreed at journal 19 Jan, 2026 Reviewers invited by journal 09 Dec, 2025 Editor invited by journal 27 Nov, 2025 Editor assigned by journal 26 Nov, 2025 Submission checks completed at journal 26 Nov, 2025 First submitted to journal 21 Nov, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8175303","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":559332777,"identity":"15b5615f-0cce-4ff4-a0d7-77a6be79a358","order_by":0,"name":"Wen Qiu","email":"","orcid":"","institution":"Guangdong Medical University","correspondingAuthor":false,"prefix":"","firstName":"Wen","middleName":"","lastName":"Qiu","suffix":""},{"id":559332778,"identity":"4a2d8dc6-5a7f-4617-957c-ba26e210a66d","order_by":1,"name":"Chunyi Liang","email":"","orcid":"","institution":"People's Hospital of Yangjiang","correspondingAuthor":false,"prefix":"","firstName":"Chunyi","middleName":"","lastName":"Liang","suffix":""},{"id":559332780,"identity":"9584cdc0-690e-497a-aef6-875a0306b4c3","order_by":2,"name":"Wenxuan Luo","email":"","orcid":"","institution":"Affiliated Hospital of Guangdong Medical University","correspondingAuthor":false,"prefix":"","firstName":"Wenxuan","middleName":"","lastName":"Luo","suffix":""},{"id":559332782,"identity":"0ab2c9c6-0fd7-4b06-88e2-5c202a069e17","order_by":3,"name":"Jinxiu Lin","email":"","orcid":"","institution":"People's Hospital of Yangjiang","correspondingAuthor":false,"prefix":"","firstName":"Jinxiu","middleName":"","lastName":"Lin","suffix":""},{"id":559332783,"identity":"518922e9-ba70-49e6-837b-5a6e0bda5d9d","order_by":4,"name":"Jianli Cao","email":"","orcid":"","institution":"People's Hospital of Yangjiang","correspondingAuthor":false,"prefix":"","firstName":"Jianli","middleName":"","lastName":"Cao","suffix":""},{"id":559332784,"identity":"e782156b-6320-4eb8-a4ba-77b69eb7c853","order_by":5,"name":"Peng Peng","email":"","orcid":"","institution":"People's Hospital of Yangjiang","correspondingAuthor":false,"prefix":"","firstName":"Peng","middleName":"","lastName":"Peng","suffix":""},{"id":559332785,"identity":"ab48fc61-96db-4007-a015-559a8fcf0479","order_by":6,"name":"Yeheng Zhou","email":"","orcid":"","institution":"People's Hospital of Yangjiang","correspondingAuthor":false,"prefix":"","firstName":"Yeheng","middleName":"","lastName":"Zhou","suffix":""},{"id":559332786,"identity":"41cfc669-94ec-4d7d-82af-2d3f0db5d62d","order_by":7,"name":"Yifan Hu","email":"","orcid":"","institution":"Central Research Institute, United Imaging Healthcare","correspondingAuthor":false,"prefix":"","firstName":"Yifan","middleName":"","lastName":"Hu","suffix":""},{"id":559332787,"identity":"0a415854-e391-4c60-a8e4-53dce5aef973","order_by":8,"name":"Renzheng Chen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAw0lEQVRIiWNgGAWjYDACCTBpw8BwAESzEa8ljXQth0nQwj+7+djjij/n7fqunTFg+FB2GCjSQMCSO8fSDc/w3E6eeTvHgHHGucNAkQP4tRhI5JhJNkjcTjYAamHmbTsMFEkgpCX/m2SDwTmIlr/Eaclhk2xIOGAH1sJIjBaJG2lAhx1ITpC8nVZwsOdcOo/EDQJa+GckP5Ns+GNnz3c7eeODH2XWcvwzCGiBgcQGBkjU8BCnHgjsiVY5CkbBKBgFIw8AAGmjQ1Zhh9QZAAAAAElFTkSuQmCC","orcid":"","institution":"Guangdong Medical University","correspondingAuthor":true,"prefix":"","firstName":"Renzheng","middleName":"","lastName":"Chen","suffix":""}],"badges":[],"createdAt":"2025-11-21 16:23:07","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8175303/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8175303/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12880-026-02299-y","type":"published","date":"2026-04-10T15:57:19+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":98348872,"identity":"589b4f24-3d42-4859-a852-0346f7a57779","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"jpg","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":100055,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/51bedc44aadf8aa34016ea38.jpg"},{"id":98439621,"identity":"ae790395-778c-41e9-8193-ea3c9b609ebe","added_by":"auto","created_at":"2025-12-17 17:02:13","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":32024,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscriptfile.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/8d4dee3ac91f035cb7ecea11.docx"},{"id":98438557,"identity":"a4a36499-1d83-4385-8897-e3d9c222955d","added_by":"auto","created_at":"2025-12-17 16:59:32","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13989,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/9dc4b3cdae3132ac3e403165.docx"},{"id":98438607,"identity":"ae5425e8-f01c-42e0-bb98-24e905f2f475","added_by":"auto","created_at":"2025-12-17 16:59:38","extension":"jpg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119441,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/0dab1c97d03aba55ead293c0.jpg"},{"id":98439284,"identity":"46898f1c-788a-4e34-8379-08e81b7a5d51","added_by":"auto","created_at":"2025-12-17 17:01:32","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12163,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/1cc33f3afbf9f5fbbd8a376a.docx"},{"id":98348887,"identity":"11aeded8-a6a8-4d38-b700-d05884eb00b0","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"jpg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":136453,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/b48dc2de06d0c466af031bf0.jpg"},{"id":98439820,"identity":"fe1fb9d3-c12d-4d17-9ea9-4060fda3cf75","added_by":"auto","created_at":"2025-12-17 17:02:58","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12781,"visible":true,"origin":"","legend":"","description":"","filename":"Table3.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/07b6c441280e9b733f551c0e.docx"},{"id":98438877,"identity":"153c8ca2-45ed-4764-aee3-3871ee07e697","added_by":"auto","created_at":"2025-12-17 17:00:20","extension":"jpg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":81421,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/a4e4fd9bebea9482fc432e3f.jpg"},{"id":98437915,"identity":"2db3475b-3242-4593-b9be-6a4b758a0309","added_by":"auto","created_at":"2025-12-17 16:58:13","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13638,"visible":true,"origin":"","legend":"","description":"","filename":"Table4.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/c12c04168f9ad84d8197b76b.docx"},{"id":98439209,"identity":"5d137555-8407-4886-b760-b06c554394a9","added_by":"auto","created_at":"2025-12-17 17:01:28","extension":"jpg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":149312,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/ecdb65c32750ade6d2f734bc.jpg"},{"id":98348895,"identity":"0de26649-ca22-492a-b279-069420031382","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"docx","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11288,"visible":true,"origin":"","legend":"","description":"","filename":"Table5.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/e0d9af69b05e68e317e986e2.docx"},{"id":98438893,"identity":"90f177c2-70d8-4c4d-b764-bacdc69a9a5c","added_by":"auto","created_at":"2025-12-17 17:00:21","extension":"docx","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1842593,"visible":true,"origin":"","legend":"","description":"","filename":"Figures.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/68f21357b530d67ff717cd7c.docx"},{"id":98439365,"identity":"d9be89bf-342b-4525-8370-0ef5d9cfce81","added_by":"auto","created_at":"2025-12-17 17:01:41","extension":"json","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11179,"visible":true,"origin":"","legend":"","description":"","filename":"f0a8bb51124e466db15e4ec2958097e9.json","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/404a6a0c4b679232ab53ffc9.json"},{"id":98437873,"identity":"da64089b-d7d5-4481-8783-861c9f0618cd","added_by":"auto","created_at":"2025-12-17 16:58:12","extension":"docx","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4852988,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFiles.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/ba93dbc9c80de7d9f895e656.docx"},{"id":98348894,"identity":"18d86cbf-caaa-4788-94c8-f36a7e13b56b","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"xml","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":116522,"visible":true,"origin":"","legend":"","description":"","filename":"f0a8bb51124e466db15e4ec2958097e91enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/f76d31e324e29dd9a24b976f.xml"},{"id":98439800,"identity":"2bd999a6-b956-4af4-a3f9-49ae683e77c6","added_by":"auto","created_at":"2025-12-17 17:02:57","extension":"jpg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":100055,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/f0109aa7bee5dbac8cadce52.jpg"},{"id":98439180,"identity":"55357d5d-cb65-4bf4-8dde-98f2238b183c","added_by":"auto","created_at":"2025-12-17 17:01:25","extension":"jpg","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119441,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/159f4f9f7794af51e3316d88.jpg"},{"id":98437650,"identity":"d51a4942-9c33-47cf-969f-ebe4efc48e24","added_by":"auto","created_at":"2025-12-17 16:57:32","extension":"jpg","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":136453,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/fba76182b88532e3ba0416d1.jpg"},{"id":98348896,"identity":"9a79038d-a881-4627-9712-7cec529156e0","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"jpg","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":81421,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/cbb46660b9c525bdb03923fd.jpg"},{"id":98439227,"identity":"d1aee29c-795c-436f-9bc2-0360a51505ad","added_by":"auto","created_at":"2025-12-17 17:01:28","extension":"jpg","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":149312,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/f25939a47c5195351cee85ed.jpg"},{"id":98348899,"identity":"64d5f088-b653-4141-a57f-be37ebcb4022","added_by":"auto","created_at":"2025-12-16 19:55:15","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":381622,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/10cf5b99b8f242be6a71c623.png"},{"id":98439573,"identity":"3b759e2c-f165-4cb4-9323-bc6be3bbe090","added_by":"auto","created_at":"2025-12-17 17:02:08","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":917816,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/88f864b38b5bfe1a2d19cbb5.png"},{"id":98438035,"identity":"6ac8189d-5a93-43d6-8b18-b9433065feb0","added_by":"auto","created_at":"2025-12-17 16:58:32","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":158015,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/46791ea0215f0eb129782eee.png"},{"id":98348901,"identity":"b35569cf-5afa-4ee6-bdf3-bcf7f7c18f2c","added_by":"auto","created_at":"2025-12-16 19:55:15","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":81861,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/821ec5a36ea43282a6bdf934.png"},{"id":98348898,"identity":"34288582-6ce9-4ccc-80c1-46860d06189e","added_by":"auto","created_at":"2025-12-16 19:55:15","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":287251,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/aedad7a8495681865388acf6.png"},{"id":98439202,"identity":"94476277-d9e0-45d4-ad8a-6cb01aefe426","added_by":"auto","created_at":"2025-12-17 17:01:27","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":95135,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/d5e79976f191b5945851ed9f.png"},{"id":98439694,"identity":"a76e5475-09aa-4f73-881d-70b93ab0a870","added_by":"auto","created_at":"2025-12-17 17:02:44","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":117848,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/11d7a966e010a5fedcddf3b3.png"},{"id":98439017,"identity":"497ce3b9-729e-44dd-820d-15ff1d506461","added_by":"auto","created_at":"2025-12-17 17:00:43","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":52938,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/b466c1cb82e1114f4fe383a3.png"},{"id":98438948,"identity":"4ade48b2-61fa-48a9-9a90-dfab501d40b5","added_by":"auto","created_at":"2025-12-17 17:00:32","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28540,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/c17ef4b92955a3d1c0a3cffe.png"},{"id":98348905,"identity":"e7968e6a-9b47-4e02-88f4-53c59face0e7","added_by":"auto","created_at":"2025-12-16 19:55:15","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70438,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/569704764eab42b9c0a32bf3.png"},{"id":98439818,"identity":"bb0436fe-7c7e-4a24-9c90-f3cca5309fb0","added_by":"auto","created_at":"2025-12-17 17:02:58","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66500,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/a7caf73e2e2b3c0453006a16.png"},{"id":98438024,"identity":"c46a5dee-e551-4ac8-bcd7-47954fd35ffe","added_by":"auto","created_at":"2025-12-17 16:58:28","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":114424,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/87f376f9fc92d33aecf437a0.png"},{"id":98348913,"identity":"90f86eb7-c78e-4919-8b91-b52b79246bf1","added_by":"auto","created_at":"2025-12-16 19:55:15","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40098,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/5b8f7a18506ccfb073a3fdd9.png"},{"id":98439322,"identity":"6fc8d975-02d0-4132-b559-a68248bcd4bd","added_by":"auto","created_at":"2025-12-17 17:01:38","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":25264,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/0657dbd99b510982de949518.png"},{"id":98438919,"identity":"1fc370aa-0dce-4d2c-80ab-fde55b9bb7ad","added_by":"auto","created_at":"2025-12-17 17:00:27","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":52285,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/5e313f10abcbc97c907f3b01.png"},{"id":98348907,"identity":"ff5fbfdc-8e63-4bb2-8810-798c46598584","added_by":"auto","created_at":"2025-12-16 19:55:15","extension":"xml","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":112222,"visible":true,"origin":"","legend":"","description":"","filename":"f0a8bb51124e466db15e4ec2958097e91structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/4ae9bfbf7a8188bfdd547648.xml"},{"id":98437662,"identity":"4fa9c2ac-c0b4-47f1-8778-aac06da12635","added_by":"auto","created_at":"2025-12-17 16:57:32","extension":"html","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125751,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/fe19d26e985275f4a1c1d750.html"},{"id":98348875,"identity":"fef9b8cd-dd29-49b8-89d0-7a2cdd453b90","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":100055,"visible":true,"origin":"","legend":"\u003cp\u003eThe workflow for this study.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/7a7902db687f4f28c8d4c25b.jpg"},{"id":98439156,"identity":"25d76be2-c813-4c82-b012-8c518023072f","added_by":"auto","created_at":"2025-12-17 17:01:18","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":119441,"visible":true,"origin":"","legend":"\u003cp\u003eIdentification of perinodular vessel counts: (A) Vessels surrounding the tumor: vessel (a, d, e) contact with the tumor on the hilar side and are identified as proximal marginal vessels, vessels (b, c) connect to the tumor only on the pleural side are thus classified as distal marginal vessels. Only proximal marginal vessels are counted to PVC. (B) Schematic representation of a subpleural tumor, in which only proximal vessels are visible. (C–F) Axial (C), coronal (D), and volumetric reconstruction images (E, F) from the same patient depict the proximal vessels adjacent to the tumor. Three proximal marginal vessels are shown in contact with the tumor (red arrows).\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/e9965300ac0b6c6fdb2e9e9e.jpg"},{"id":98437604,"identity":"dd1cdbc8-92f1-405e-9cc6-84fcefcbce00","added_by":"auto","created_at":"2025-12-17 16:57:30","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":136453,"visible":true,"origin":"","legend":"\u003cp\u003eExperiment results for model selection. (a) comparison between the clinical models before and after adding the key feature PVC; (b) ROC curve comparison of three ablation experimental models; (c) confusion matrices of the above four models.\u003c/p\u003e","description":"","filename":"Figure3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/e9d37f01972c82d02cc19f5c.jpg"},{"id":98348881,"identity":"e5ec0157-bec1-4e5b-838d-517f39e1af9d","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":81421,"visible":true,"origin":"","legend":"\u003cp\u003eExperiment results for pure clinical model and the CIPR model evaluation on multicenter test dataset, as well as the diagnostic performances of two senior radiologists and one junior radiologist. (a) results on internal test set (b) results on external test set.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/95bce44a7e0f00027a47fc2a.jpg"},{"id":98348877,"identity":"aa90285a-742a-42f0-acce-094f44c910be","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":149312,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualization of Interpretability Analysis for PVC.\u003c/strong\u003e (a) violin plot illustrating the distribution of the PVC feature in the training set; (b) the importance scores assigned to the PVC feature by the MRMR algorithm; (c) the importance scores assigned to the PVC feature by the Relief algorithm; (d) SHAP summary plot for the CIPR model applied to the internal independent test set. (e) the distribution of SHAP values across the test set, along with the SHAP dependence plot for the PVC feature. (f) the SHAP interaction effects between the PVC feature and the Density Uniform feature.\u003c/p\u003e","description":"","filename":"Figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/e2ea28d5b408b27119bd9121.jpg"},{"id":106808758,"identity":"8f867efc-554b-49ee-91d2-a734f236e184","added_by":"auto","created_at":"2026-04-13 16:00:53","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1413812,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/c5902cb8-04fa-4efc-98e0-7e6bbaf8dd99.pdf"},{"id":98348891,"identity":"1ddc202a-27f3-4c74-9b3a-7f0a58e8a281","added_by":"auto","created_at":"2025-12-16 19:55:14","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":4852988,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFiles.docx","url":"https://assets-eu.researchsquare.com/files/rs-8175303/v1/f8161f9a4eacec51fc92adde.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Improving Risk Stratification of Pulmonary Nodules: An Integrated Perinodular Vascular and Radiomic Model for Clinical Decision Support","fulltext":[{"header":"Key points","content":"\u003cp\u003eQuestion: To achieve the\u0026nbsp;early and accurate classification of pulmonary nodules, developing\u0026nbsp;models based on more robust perinodular nodule features is crucial.\u003c/p\u003e\n\u003cp\u003eFindings: A combined model utilizing clinical and radiomic features, particularly perinodular vascular count, enhances the accuracy and generalizability of pulmonary nodule evaluation on multicenter dataset.\u003c/p\u003e\n\u003cp\u003eClinical Relevance: This study offers a clinical decision-support tool for clinicians to reduce overtreatment by reliably ruling out malignancy in ambiguous pulmonary nodules, thus preventing unnecessary invasive procedures and improving patient care.\u003c/p\u003e"},{"header":"Introduction","content":"\u003cp\u003eLung cancer has reemerged as the most commonly diagnosed cancer worldwide, with 2.481\u0026nbsp;million new cases in 2022 accounting for 12.4% of all cancer diagnoses [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Moreover, it is projected to become the leading cause of global cancer incidence and mortality by 2050 [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Prognosis is highly stage-dependent: the five-year survival rate is 77\u0026ndash;92% for Stage I but plummets to 0\u0026ndash;36% for Stage II, underscoring the critical importance of early detection [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. However, a majority of lung cancer patients are diagnosed at advanced stages, missing the window for curative treatment.\u003c/p\u003e\u003cp\u003eA pulmonary nodule (PN) is a focal, round, high-density shadow in the lung, either solid or subsolid, with a diameter of \u0026le;\u0026thinsp;30 mm. It serves as a \"carrier\" or early manifestation of lung cancer, with the vast majority of lung cancers first detected as nodules. Nevertheless, most pulmonary nodules are not cancerous but are caused by benign conditions such as infection or inflammation. Consequently, the clinical management strategy for a pulmonary nodule depends directly on whether it is benign or malignant. The International Agency for Research on Cancer provides clear pathological criteria. For nodules pathologically identified as benign, such as preinvasive glandular lesions, active surveillance is typically recommended. In contrast, malignant nodules generally require surgical resection [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. To reduce over-treatment rates without compromising the malignant detection rate, thus allowing for effective stratified management of nodules\u0026mdash;has become a major clinical challenge.\u003c/p\u003e\u003cp\u003eThe imprecision of conventional radiological assessment in characterizing pulmonary nodules can result in both delayed treatment and unnecessary procedures. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] Radiomics and other AI approaches have shown promise in improving diagnostic accuracy by extracting high-dimensional quantitative features from medical images. [\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] These intranodular features can capture tumor heterogeneity and have proven diagnostic value. However, tumors do not exist in isolation. Evidence shows that perinodular characteristics, like changes in surrounding vasculature, are key markers of tumor aggressiveness and prognosis, highlighting a vital direction for AI-driven diagnosis. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/p\u003e\u003cp\u003ePrevious research on lung nodule diagnosis has largely examined intranodular radiomics, perinodular features, and clinical factors in isolation, with a notable scarcity of integrated models and a lack of quantitative validation for perinodular characteristics. To address this gap, our study proposes a comprehensive model that integrates these three domains to enhance diagnostic precision. Moreover, a key objective is the interpretable quantification of PVC. We anticipate that this holistic strategy will yield deeper insights into nodule behavior and directly support improved clinical decision-making.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eThis study aims to develop a machine learning model integrating multimodal features, which will subsequently be evaluated on multi-center test sets to assess its generalizability. Furthermore, a rigorous interpretability analysis was conducted to validate the importance of the PVC feature. The specific workflow is outlined in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eDataset\u003c/h2\u003e\u003cp\u003eThis retrospective study curated a primary dataset of 218 nodules from 100 patients using CT scans acquired between January 2023 and December 2024 from hospital 1. An independent internal test set of 81 nodules from 62 patients was formed from scans dated January to July 2025, and an external test set of 21 nodules from 16 patients was included from hospital 2. All cases met the following criteria: (a) availability of postoperative pathological results; (b) nodule maximum diameter\u0026thinsp;\u0026le;\u0026thinsp;30 mm; (c) CT scan resolution\u0026thinsp;\u0026le;\u0026thinsp;1 mm and slice thickness\u0026thinsp;\u0026lt;\u0026thinsp;1.5 mm; and (d) availability of essential clinical characteristics. The detailed distribution of the data is shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003epresents the distribution of important features related to the data in the training set, validation set, and two test sets.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVariable\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTraining\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eValidation\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eIndependent Test\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eExternal Test\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNo. of Lesions\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e186\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e81\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e21\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e57.01\u0026thinsp;\u0026plusmn;\u0026thinsp;12.43\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e58.06\u0026thinsp;\u0026plusmn;\u0026thinsp;12.81\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e60.22\u0026thinsp;\u0026plusmn;\u0026thinsp;11.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e59.33\u0026thinsp;\u0026plusmn;\u0026thinsp;11.99\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGender\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e122\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFemale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e64\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePVC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2.59\u0026thinsp;\u0026plusmn;\u0026thinsp;1.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2.94\u0026thinsp;\u0026plusmn;\u0026thinsp;1.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2.65\u0026thinsp;\u0026plusmn;\u0026thinsp;1.39\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3.1\u0026thinsp;\u0026plusmn;\u0026thinsp;2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHomogeneous density\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e66%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e66%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e48%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNodule type\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSolid\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePart-Solid\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e25\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGround-Glass\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e109\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e36\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRadiological signs\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLobulation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e23%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e25%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e31%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e48%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePleural indentation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e22%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e38%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e38.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e67%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVascular convergence\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e16%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e22.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e43%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVacuole\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e24%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e16%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e20%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e24%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eFeature Extraction and Selection\u003c/h3\u003e\n\u003cp\u003eThree categories of features were extracted from each pulmonary nodule for comprehensive analysis. All features included are routinely utilized as standard measures in clinical and research settings. The first category comprised 30 clinical, laboratory, and radiological features, including patient demographics (age, gender, cough, sputum, hemoptysis, fever, chest pain), medical history (smoking, diabetes, tuberculosis, cancer, family), white blood cell count, carcinoembryonic antigen level, nodule morphological characteristics(nodule type, long-axis diameter, long-axis to short-axis ratio, volume, homogeneous density, boundary clarity, morphological regularity), imaging signs (lobulation, spiculation, pleural indentation, vascular convergence, vacuole sign, calcification, cavity), and key imaging signs\u0026mdash;with perinodular vessel count (PVC) being a focus of this study. [\u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] For the PVC feature, we established a criterion in Fig.\u0026nbsp;2 and performed the annotation based on this standard. Clinical and laboratory features were obtained through retrospective review of patient outpatient and laboratory records. Each imaging sign for every nodule was independently annotated by two resident physicians, with any discrepancies adjudicated by a senior attending radiologist.\u003c/p\u003e\u003cp\u003eFigure 2. Identification of perinodular vessel counts: (A) Vessels surrounding the tumor: vessel (a, d, e) contact with the tumor on the hilar side and are identified as proximal marginal vessels, vessels (b, c) connect to the tumor only on the pleural side are thus classified as distal marginal vessels. Only proximal marginal vessels are counted to PVC. (B) Schematic representation of a subpleural tumor, in which only proximal vessels are visible. (C\u0026ndash;F) Axial (C), coronal (D), and volumetric reconstruction images (E, F) from the same patient depict the proximal vessels adjacent to the tumor. Three proximal marginal vessels are shown in contact with the tumor (red arrows).\u003c/p\u003e\u003cp\u003eThe second and third categories consisted of 1,927 radiomic features each, extracted from the core nodule region and a 3-voxel dilated perinodular region, respectively, where each region of interest (nodule) was meticulously delineated. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] These radiomic features encompassed intensity, texture, and shape information, and were automatically generated using the United Imaging Research Platform following semi-automatic segmentation.\u003c/p\u003e\u003cp\u003eTo address high dimensionality and potential redundancy in the initial feature set (~\u0026thinsp;4,000 features), a two-stage filter-based selection strategy was implemented. The Maximum Relevance Minimum Redundancy (MRMR) [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] algorithm was first applied for global redundancy reduction, prioritizing features with high relevance to malignancy and low inter-feature redundancy based on mutual information. This was followed by the Relief algorithm [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] to refine local discriminatory power by evaluating feature effectiveness in distinguishing benign and malignant cases within local instance neighborhoods. This combined approach effectively balanced redundancy control with classification boundary optimization while preserving predictive performance and interpretability.\u003c/p\u003e\u003cp\u003eIn subsequent modeling, feature selection was adapted to the input type. For clinical-only models, Relief was used to select the top 30% of features (approximately 9 features). For combined clinical-radiomic models, MRMR first reduced the feature set to the top 20%, from which Relief then selected a final set of 18 features\u0026mdash;double the number used in clinical-only models\u0026mdash;to fully leverage multimodal information.\u003c/p\u003e\n\u003ch3\u003eClassification\u003c/h3\u003e\n\u003cp\u003eThis study was designed with a systematic experimental protocol to comprehensively evaluate the performance, robustness, and practical potential of the proposed multimodal classification model in real-world clinical settings. The core objectives included: (1) validating the necessity and contribution of the key clinical feature PVC to model performance; (2) elucidating the synergistic effects and individual value of features derived from different sources (clinical features, intranodular radiomic features, and perinodular radiomic features); and (3) assessing the model\u0026rsquo;s advantage over pure human diagnosis and its potential to reduce the rate of overtreatment. All experiments were conducted using rigorously partitioned training, validation, and independent test sets.\u003c/p\u003e\u003cp\u003eFirst, to accurately evaluate the role of perinodular vessel count in the classification task, an ablation study at the feature level was designed. By comparing models trained on the full feature set and model intentionally ablated by excluding the vessel count feature, the necessity and independent contribution of this feature to overall performance were quantified.\u003c/p\u003e\u003cp\u003eSecond, to dissect the synergistic mechanisms among features from different sources, a progressive feature integration strategy was adopted. A baseline model relying solely on structured clinical features was initially established. Intranodular radiomic features extracted from the target region of interest (ROI) were then incrementally incorporated to evaluate their added value beyond clinical information, which is the Clinic_Intranodular_Radiomics (CIR) model. Finally, perinodular radiomic features were encoded and integrated to construct a comprehensive multimodal fusion model, which is the Clinic_Intranodular_Perinodular_Radiomics (CIPR) model, thoroughly validating the complementarity and integrative efficacy of cross-source features.\u003c/p\u003e\u003cp\u003eLastly, a controlled experiment involving specialized physicians was conducted to quantify the auxiliary value of the model in real clinical practice. The physician group performed independent diagnoses without assistance and recorded their conclusions and the model group generated predictions and confidence scores autonomously. Comparisons among these two groups were made along two dimensions\u0026mdash;diagnostic sensitivity and specificity\u0026mdash;to elucidate the practical advantages of human-AI collaborative decision-making.\u003c/p\u003e\u003cp\u003eThe primary dataset was randomly partitioned into a training set and an internal test set in an 85:15 ratio. Both the independent and external test sets were strictly designed to include patients and nodules distinct from those in the training set, ensuring robust evaluation of model generalizability. For model training, all experiments were conducted using a bagging ensemble of decision trees [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] implemented on the uAI Research Portal (uAI Research Portal, Version: 20240130) [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Specifically, the Scikit-Learn library, an open-source machine learning library, was utilized to build the model. The number of trees was set to half the number of features to avoid overfitting, and the tree depth was set to the platform\u0026rsquo;s default value.\u003c/p\u003e\n\u003ch3\u003eFeature Interpretability Analysis\u003c/h3\u003e\n\u003cp\u003eA multi-level analysis was employed to interpret the selected features. First, PVC\u0026rsquo;s importance was objectively assessed using MRMR (reflecting feature relevance and redundancy) and Relief (confirming classification utility) algorithms. Second, SHAP analysis globally quantified the contribution of PVC to model predictions, illustrating how its variation affects output probabilities. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/p\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eModel predictions were evaluated using confusion matrix-derived performance metrics, including Accuracy, Sensitivity, Specificity, Youden Index, and the F1-Score, to comprehensively assess classification efficacy. Given the limited sample size, bootstrap resampling [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] (100 iterations) was applied on results of each test set to estimate confidence intervals for each metric. Inter-group comparisons were performed using Student\u0026rsquo;s t-test [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], with a significance threshold set at α\u0026thinsp;=\u0026thinsp;0.05. Key results are reported as mean and confidence interval. All analyses were conducted on both validation and test sets, including the internal validation set, the internal independent test set, and the external test set, to ensure unbiased evaluation.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003eClassification Results\u003c/h2\u003e\u003cp\u003eAs illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a) and Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, an ablation experiment focusing on the perinodular vessel count (PVC) demonstrated that incorporating this feature into the feature set\u0026mdash;followed by feature selection\u0026mdash;significantly enhanced the AUC to 0.8178, markedly outperforming the model trained without PVC. Correspondingly, sensitivity, specificity, and F1-score improved substantially from 0.6316, 0.6154, and 0.6667 to 0.8421, 0.7692, and 0.8421, respectively. These results indicate that the inclusion of PVC significantly boosts the diagnostic performance of conventional clinical features in distinguishing benign from malignant lesions.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eModel Comparison of the Malignant-Benign Classification Ability of PVC Features\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSen\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSpe\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eAcc\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003ePre\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClinic\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.8178\u003c/p\u003e\u003cp\u003e(0.6417, 0.9676)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.8421\u003c/p\u003e\u003cp\u003e(0.6581, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.7692\u003c/p\u003e\u003cp\u003e(0.5158, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.8125\u003c/p\u003e\u003cp\u003e(0.6562, 0.9375)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.6113\u003c/p\u003e\u003cp\u003e(0.3133, 0.8693)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.8421\u003c/p\u003e\u003cp\u003e(0.6667, 0.9511)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCIPR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.7389\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.5448, 0.9076)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.6316\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.4148, 0.8534)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.6154\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.3343, 0.8182)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.6250\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.4523, 0.7812)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.2470\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(-0.0873, 0.5918)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.6667\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.4823, 0.8392)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u003cb\u003eNote\u003c/b\u003e: Significance levels are denoted as follows: *p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, **p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ***p\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn the multi-source feature ablation experiments, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e(b) and Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the integration of radiomic features led to a notable increase in AUC from 0.8178 to 0.8704. Furthermore, Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e reveal that the simultaneous incorporation of both intranodular and perinodular radiomic features yielded significant improvements in AUC and specificity compared to using intranodular features alone. We also compared clinic model against CIPR model on two independent test sets, demonstrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The CIPR model consistently outperformed the clinic model across both datasets. Notably, on the Yangjiang independent test set, the CIPR model achieved significantly higher values across all classification metrics. Specifically, it exhibited superior sensitivity and F1-scores on both datasets, while matching or exceeding the clinic model in other performance indicators. These findings underscore the substantial enhancement in discriminative efficacy brought by radiomic features, particularly those related to perinodular vascular patterns.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eAblation study of clinical features including PVC, intranodular and perinodular radiomics\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSen\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSpe\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eAcc\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eYouden\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCIPR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.8704\u003c/p\u003e\u003cp\u003e(0.6417, 0.9676)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.8421\u003c/p\u003e\u003cp\u003e(0.6581, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.7692\u003c/p\u003e\u003cp\u003e(0.5158, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.8125\u003c/p\u003e\u003cp\u003e(0.6562, 0.9375)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.6113\u003c/p\u003e\u003cp\u003e(0.3133, 0.8693)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.8421\u003c/p\u003e\u003cp\u003e(0.6667, 0.9511)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCIR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.8158\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.6397, 0.9543)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.8947\u003c/p\u003e\u003cp\u003e(0.7099, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.6923\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.5158, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.8125\u003c/p\u003e\u003cp\u003e(0.6398, 0.9375)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.5870\u003c/p\u003e\u003cp\u003e(0.2297, 0.8745)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.8500\u003c/p\u003e\u003cp\u003e(0.6766, 0.9545)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClinic\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.8178\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.6417, 0.9676)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.8421\u003c/p\u003e\u003cp\u003e(0.6581, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.7692\u003c/p\u003e\u003cp\u003e(0.5158, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.8125\u003c/p\u003e\u003cp\u003e(0.6562, 0.9375)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.6113\u003c/p\u003e\u003cp\u003e(0.3133, 0.8693)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.8421\u003c/p\u003e\u003cp\u003e(0.6667, 0.9511)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u003cb\u003eNote\u003c/b\u003e: Significance levels are denoted as follows: *p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, **p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ***p\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of classification results between CIPR and clinic model on independent and external test dataset\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSen\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSpe\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAcc\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eYouden\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eIndependent\u003c/p\u003e\u003cp\u003eTesting\u003c/p\u003e\u003cp\u003eDataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCIPR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.8225\u003c/p\u003e\u003cp\u003e(0.7298, 0.9168)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8824\u003c/p\u003e\u003cp\u003e(0.8059, 0.9622)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.7000\u003c/p\u003e\u003cp\u003e(0.5601, 0.8571)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.8148\u003c/p\u003e\u003cp\u003e(0.7343, 0.8889)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.5824\u003c/p\u003e\u003cp\u003e(0.4088, 0.7563)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.8571\u003c/p\u003e\u003cp\u003e(0.7770, 0.9167)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eClinic\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.8007\u003csup\u003e**\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.6946, 0.9025)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8235\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.7315, 0.9262)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.6667\u003csup\u003e**\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.4835, 0.8458)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.7654\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.6725, 0.8519)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.4902\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.2728, 0.6929)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.8155\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.7368, 0.8879)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eExternal\u003c/p\u003e\u003cp\u003eDataset\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCIPR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.7937\u003c/p\u003e\u003cp\u003e(0.4234, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8750\u003c/p\u003e\u003cp\u003e(0.6750, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.8\u003c/p\u003e\u003cp\u003e(0.3786, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.8571\u003c/p\u003e\u003cp\u003e(0.6667, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.6750\u003c/p\u003e\u003cp\u003e(0.2181, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.9032\u003c/p\u003e\u003cp\u003e(0.7586, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eClinic\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.7562\u003c/p\u003e\u003cp\u003e(0.3438, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8125\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.5714, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.8\u003c/p\u003e\u003cp\u003e(0.3786, 1.0000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.8095\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.6190, 0.9524)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.6125\u003csup\u003e*\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.1389, 0.9460)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.8667\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e(0.6804, 0.9722)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eNote\u003c/b\u003e: Significance levels are denoted as follows: *p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, **p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ***p\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e further demonstrates that the model overall matched or exceeded the diagnostic performance of two senior radiologists and one junior radiologist on the independent internal test set. Although two radiologists achieved specificities above 0.8, their sensitivities were both below 0.7, considerably lower than the model\u0026rsquo;s sensitivity of 0.8824. Moreover, in terms of the more balanced metric\u0026mdash;Youden Index\u0026mdash;the model attained a value of 0.5824 on the internal test set, outperforming the three radiologists\u0026rsquo; scores of 0.4667, 0.5627, and 0.4608, respectively. On the external test set, the model achieved a Youden Index of 0.6750, significantly surpassing the radiologists\u0026rsquo; results of 0.4875, 0.0125, and 0.225. These results indicate that, compared with human evaluators, the model exhibits superior balance between sensitivity and specificity, thereby assisting clinicians in avoiding overdiagnosis while maintaining a high malignant detection rate.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eInterpretability Analysis of PVC\u003c/h3\u003e\n\u003cp\u003eOur analysis reaffirms that the PVC feature, along with other medical imaging-related characteristics, plays a pivotal role in significantly enhancing the diagnostic performance of the model. Subsequent interpretability analyses\u0026mdash;including feature importance metrics and SHAP analysis\u0026mdash;consistently underscore the critical importance of the PVC feature.\u003c/p\u003e\u003cp\u003eFirst, examination of feature distributions within the training set reveals a statistically significant disparity in PVC values between benign and malignant cases, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a). During feature selection, the MRMR algorithm assigned the highest score to the PVC feature, indicating that it achieves an optimal balance between maximal relevance to the target variable and minimal redundancy with other features (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b)). In other words, the algorithm identifies PVC as providing the greatest information gain with the least overlap. In a subsequent feature ranking using the Relief algorithm, PVC was considered as the second-highest importance score, immediately following \u0026ldquo;Homogeneous Density\u0026rdquo;, further corroborating its discriminative power in classifying benign and malignant nodules (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e(c)).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSHAP analysis likewise highlighted the prominence of the PVC feature. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e(d), PVC attained the second-highest mean |SHAP value|, confirming its role as one of the most influential predictors at both local and global interpretation levels. This offers robust and explainable evidence that the model heavily relies on this feature during decision-making. Furthermore, the SHAP dependence plot [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e(e) demonstrates that higher PVC values correlate with increased model tendency toward predicting malignancy\u0026mdash;a finding consistent with clinical intuition. Additionally, Homogeneous Density also emerged as a major contributor. The SHAP interaction plot [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e(f) reveals notable interaction effects between these two features: particularly in cases with non-uniform density, PVC serves as an even more decisive factor. In contrast, when nodule density is uniform, their joint influence approaches zero, implying that additional features are required to guide the model\u0026rsquo;s decision.\u003c/p\u003e\u003cp\u003eIn conclusion, the PVC feature constitutes a cornerstone of the model\u0026rsquo;s ability to differentiate between benign and malignant nodules. As an interpretable decision-support tool, it holds considerable promise for integration into clinical workflows, thereby aiding diagnosticians in making more informed and reliable decisions.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eEarly and accurate classification of pulmonary nodules based on conventional CT features is clinically challenging due to nonspecific imaging appearances, technical variability, and limited observer agreement. Current criteria\u0026mdash;such as size, density, and morphology\u0026mdash;often lead to diagnostic overlap and high false-positive rates, particularly in intermediate-sized nodules [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Furthermore, acquisition parameters (e.g., slice thickness, dose) significantly affect diagnostic outputs [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], and inter-observer consistency remains moderate (κ\u0026thinsp;=\u0026thinsp;0.454) [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. Similarly, AI models continue to produce false results and exhibit bias, restricting clinical utility [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. This study therefore seeks to identify more robust discriminative features to improve nodule assessment.\u003c/p\u003e\u003cp\u003eThis study confirms the diagnostic significance of PVC and establishes that a combined model (CIPR) incorporating clinical and radiomic features\u0026mdash;particularly those capturing perinodular vascular geometry\u0026mdash;achieves more accurate and generalizable performance than clinical models alone. Furthermore, the AI approach provided more consistent and objective assessments compared to physician judgments, highlighting its potential as a decision-support tool to mitigate overtreatment in cases of diagnostic uncertainty. Clinically, our model shows strong potential to reduce over-treatment. According to all datasets, nodules with \u0026ge;\u0026thinsp;4 perinodular vessels were rarely benign (4/78), indicating low over-treatment risk. More importantly, in cases with fewer vessels\u0026mdash;where visual assessment is difficult\u0026mdash;the model provided high specificity (\u0026gt;\u0026thinsp;70%) without compromising sensitivity (\u0026ge;\u0026thinsp;70%), as shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, enabling more confident rule-out of malignancy and avoiding unnecessary procedures in ambiguous cases.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eCIPR model classification results on nodules with PVC\u0026thinsp;\u0026lt;\u0026thinsp;4\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDataset\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAmount of PVC\u0026thinsp;\u0026lt;\u0026thinsp;4\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eValidation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7272\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.8333\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIndependent Test\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7931\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.7\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eExternal Test\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7272\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.8\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThis study has several limitations. The sample size, particularly the small number of benign cases (n\u0026thinsp;=\u0026thinsp;5) in the external test set, may affect the robustness of the evaluation. Furthermore, interpreter subjectivity affects vessel count assessment due to differences in physician qualifications and image quality, potentially influencing model predictions. Future work should focus on establishing a unified vessel counting protocol, elucidating the pathological basis of perinodular vascular features, and developing automated quantification methods to improve objectivity and clinical integration.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThanks to \u003cstrong\u003eCentral Research Institute, United Imaging Healthcare\u003c/strong\u003e for their technical support. We also thank Guangdong Medical University Affiliated Hospital for providing the external test dataset.We also extend our heartfelt appreciation to the anonymous reviewers, whose diligent efforts significantly enhanced the quality of this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;Author contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eW Q conceptualized the study, curated the data, and conducted formal analysis along. W Q, C-Y L, and W-X L were responsible for the investigation. The methodology was developed by W Q and Y-F H. Project administration was managed by W Q. Resources were provided by W Q, W-X L, J-X L, J-L C, P Pand Y-H Z. Software development was handled by Y-F H. Supervision and validation were conducted by Y-F H and R-Z C. Visualization was carried out by W Q. W Q and Y-F H prepared the original draft, while Y-F H and R-Z C reviewed and edited the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study has received funding by the Yangjiang Science and Technology Bureau,China(Grant No. SF2025055).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee of Yangjiang People\u0026rsquo;s Hospital (Approval No.KY-2025-42-1). The Committee waived the requirement for patient consent for this investigation. The authors declare that the research involving human data complies with the Helsinki Declaration.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Publish declaration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability declaration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data and tables included in this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflicts of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74(3):229\u0026ndash;63. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3322/caac.21834\u003c/span\u003e\u003cspan address=\"10.3322/caac.21834\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang C, Shao J, Song L, Ren P, Liu D, Li W. Persistent increase and improved survival of stage I lung cancer based on a large-scale real-world sample of 26,226 cases. Chin Med J (Engl). 2023;136(16):1937\u0026ndash;48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/CM9.0000000000002729\u003c/span\u003e\u003cspan address=\"10.1097/CM9.0000000000002729\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBizuayehu HM, Ahmed KY, Kibret GD, et al. Global Disparities of Cancer and Its Projected Burden in 2050. JAMA Netw Open. 2024;7(11):e2443198. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1001/jamanetworkopen.2024.43198\u003c/span\u003e\u003cspan address=\"10.1001/jamanetworkopen.2024.43198\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2024 Nov 4.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChinese Thoracic Society. Chinese expert consensus on diagnosis of early lung cancer (2023 Edition). Zhonghua Jie He He Hu Xi Za Zhi. 2023; 46:1\u0026ndash;18. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3760/cma.j.cn112147-20220712-00592\u003c/span\u003e\u003cspan address=\"10.3760/cma.j.cn112147-20220712-00592\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTsao MS, Nicholson AG, Maleszewski JJ, Marx A, Travis WD. Introduction to 2021 WHO Classification of Thoracic Tumors. J Thorac Oncol. 2022;17(1):e1\u0026ndash;4. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jtho.2021.09.017\u003c/span\u003e\u003cspan address=\"10.1016/j.jtho.2021.09.017\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang W, Yang B, Wu H, et al. Auxiliary Diagnosis of Pulmonary Nodules' Benignancy and Malignancy Based on Machine Learning: A Retrospective Study. J Multidiscip Healthc. 2025;18:3735\u0026ndash;48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2147/JMDH.S518166\u003c/span\u003e\u003cspan address=\"10.2147/JMDH.S518166\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2025 Jun 27.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen D, Yang L, Zhang W, Shen J, Van Schil PEY, Divisi D, Seetharamu N, Gu J. Prevalence and management of pulmonary nodules: a systematic review and meta-analysis. J Thorac Dis. 2024;16(7):4619\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21037/jtd-24-874\u003c/span\u003e\u003cspan address=\"10.21037/jtd-24-874\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Epub 2024 Jul 24. PMID: 39144359; PMCID: PMC11320231.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAyasa Y, Alajrami D, Idkedek M, Tahayneh K, Akar FA. The Impact of Artificial Intelligence on Lung Cancer Diagnosis and Personalized Treatment. Int J Mol Sci. 2025;26(17):8472. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/ijms26178472\u003c/span\u003e\u003cspan address=\"10.3390/ijms26178472\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2025 Aug 31.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang C, Shao J, He Y, et al. Data-driven risk stratification and precision management of pulmonary nodules detected on chest computed tomography. Nat Med. 2024;30:3184\u0026ndash;95. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41591-024-03211-3\u003c/span\u003e\u003cspan address=\"10.1038/s41591-024-03211-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang Y, Wang P, Yu C, et al. Application of artificial intelligence medical imaging aided diagnosis system in the diagnosis of pulmonary nodules. BMC Med Inf Decis Mak. 2025;25(185). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12911-025-03009-4\u003c/span\u003e\u003cspan address=\"10.1186/s12911-025-03009-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang XC, Lv FJ, Fu BJ, Liang ZR, Chu ZG. Significance of marginal vessels in differentiating peripheral small-cell lung cancer and benign lung tumor. Acta Radiol. 2023;64(9):2526\u0026ndash;34. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/02841851231188060\u003c/span\u003e\u003cspan address=\"10.1177/02841851231188060\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang W, Qi X, He Y, et al. Radiomics of Vascular Structures in Pulmonary Ground-Glass Nodules: A Predictor of Invasiveness: Radiomics of Vascular Structures in GGNs for Tumor Invasiveness Prediction. Curr Med Imaging. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2174/0115734056385352250410053810\u003c/span\u003e\u003cspan address=\"10.2174/0115734056385352250410053810\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMacMahon H, Naidich DP, Goo JM, et al. Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017. Radiology. 2017;284(1):228\u0026ndash;43. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/radiol.2017161659\u003c/span\u003e\u003cspan address=\"10.1148/radiol.2017161659\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBaeza S, Gil D, Sanchez C, et al. Radiomics and Clinical Data for the Diagnosis of Incidental Pulmonary Nodules and Lung Cancer Screening: Radiolung Integrative Predictive Model. Arch Bronconeumol. 2024;60(Suppl 2):S22\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.arbres.2024.05.027\u003c/span\u003e\u003cspan address=\"10.1016/j.arbres.2024.05.027\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZwirewich CV, Vedal S, Miller RR, M\u0026uuml;ller NL. Solitary pulmonary nodule: high-resolution CT and radiologic-pathologic correlation. Radiology. 1991;179(2):469\u0026ndash;76. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/radiology.179.2.2014294\u003c/span\u003e\u003cspan address=\"10.1148/radiology.179.2.2014294\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSelvam M, Chandrasekharan A, Sadanandan A, et al. Radiomics as a non-invasive adjunct to Chest CT in distinguishing benign and malignant lung nodules. Sci Rep. 2023;13(19062). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-023-46391-7\u003c/span\u003e\u003cspan address=\"10.1038/s41598-023-46391-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMasquelin AH, Alshaabi T, Cheney N, Est\u0026eacute;par RSJ, Bates JHT, Kinsey CM. Perinodular Parenchymal Features Improve Indeterminate Lung Nodule Classification. Acad Radiol. 2023;30(6):1073\u0026ndash;80. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.acra.2022.07.001\u003c/span\u003e\u003cspan address=\"10.1016/j.acra.2022.07.001\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePeng H, Long F, Ding C. Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell. 2005;27(8):1226\u0026ndash;38.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim K, Rendell LA. Morgan Kaufmann, Los Altos, CA,. A practical approach to feature selection, in: Proceedings 9th international Conference on Machine Learning, Aberdeen, Scotland (1992).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAltman N, Krzywinski M. Ensemble methods: bagging and random forests. Nat Methods. 2017;14:933\u0026ndash;4. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/nmeth.4438\u003c/span\u003e\u003cspan address=\"10.1038/nmeth.4438\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePlaia A, Buscemi S, F\u0026uuml;rnkranz J, et al. Comparing Boosting and Bagging for Decision Trees of Rankings. J Classif. 2022;39:78\u0026ndash;99. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00357-021-09397-2\u003c/span\u003e\u003cspan address=\"10.1007/s00357-021-09397-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWu J, Xia Y, Wang X, et al. uRP: An integrated research platform for one-stop analysis of medical images. Front Radiol. 2023;3:1153784. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fradi.2023.1153784\u003c/span\u003e\u003cspan address=\"10.3389/fradi.2023.1153784\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2023 Apr 18.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLundberg S, Lee S. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17). 2017; Curran Associates Inc., Red Hook, NY, USA, 4768\u0026ndash;4777.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePonce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024;17(11):e70056. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/cts.70056\u003c/span\u003e\u003cspan address=\"10.1111/cts.70056\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 39463176; PMCID: PMC11513550.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEfron B. Nonparametric estimates of standard error: The jackknife, the bootstrap and other methods. Biometrika. 1981;68:589\u0026ndash;99.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCho K. Performance assessment through bootstrap. IEEE Trans Pattern Anal Mach Intell. 1997;19:1185\u0026ndash;98.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFisher RA. Applications of \u0026lsquo;Student\u0026rsquo;s\u0026rsquo; distribution. Metron. 1925;5:90\u0026ndash;104.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHajjar M, Aldabbagh G, Albaradei S. Interpretable Multi-Cancer Early Detection Using SHAP-Based Machine Learning on Tumor-Educated Platelet RNA. Diagnostics (Basel). 2025;15(17):2216. Published 2025 Sep 1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/diagnostics15172216\u003c/span\u003e\u003cspan address=\"10.3390/diagnostics15172216\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFumagalli F, Muschalik M, Kolpaczki P, H\u0026uuml;llermeier E, Hammer BE. SHAP-IQ: Unified approximation of any-order shapley interactions. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023), 2023.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYao Y, Wang X, Guan J et al. Metabolomic differentiation of benign vs malignant pulmonary nodules with high specificity via high-resolution mass spectrometry analysis of patient sera. Nat Commun. 2023;14(1):2339. Published 2023 Apr 24. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-023-37875-1\u003c/span\u003e\u003cspan address=\"10.1038/s41467-023-37875-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang C, Shao J, He Y, et al. Data-driven risk stratification and precision management of pulmonary nodules detected on chest computed tomography. Nat Med. 2024;30(11):3184\u0026ndash;95. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41591-024-03211-3\u003c/span\u003e\u003cspan address=\"10.1038/s41591-024-03211-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMartini K, Ottilinger T, Serrallach B, et al. Lung cancer screening with submillisievert chest CT: Potential pitfalls of pulmonary findings in different readers with various experience levels. Eur J Radiol. 2019;121:108720. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.ejrad.2019.108720\u003c/span\u003e\u003cspan address=\"10.1016/j.ejrad.2019.108720\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao M, Xue G, He B, et al. Integrated multiomics signatures to optimize the accurate diagnosis of lung cancer. Nat Commun. 2025;16(1):84. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-024-55594-z\u003c/span\u003e\u003cspan address=\"10.1038/s41467-024-55594-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSourlos N, Wang J, Nagaraj Y, van Ooijen P, Vliegenthart R. Possible Bias in Supervised Deep Learning Algorithms for CT Lung Nodule Detection and Classification. Cancers (Basel). 2022;14(16):3867. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/cancers14163867\u003c/span\u003e\u003cspan address=\"10.3390/cancers14163867\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Published 2022 Aug 10.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Computed tomography, Benign-Malignant Classification, Pulmonary Nodules, Perinodular Vascular Count, Combined Clinical-Radiomic-Features Model, Feature Interpretability","lastPublishedDoi":"10.21203/rs.3.rs-8175303/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8175303/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eObjectives\u003c/h2\u003e\u003cp\u003eLung cancer is the leading cause of cancer-related deaths worldwide, with most patients diagnosed at advanced stages. Accurate differentiation of benign and malignant pulmonary nodules remains a major clinical challenge.\u003c/p\u003e\u003ch2\u003eMaterials and Methods\u003c/h2\u003e\u003cp\u003eWe established and validated the perinodular vessel count (PVC) as an instrumental imaging biomarker, demonstrating its significant contribution to discriminating malignant pulmonary nodules. Leveraging this finding, we constructed an integrated predictive model incorporating intranodular and perinodular radiomics, PVC, and relevant clinical variables. A two-tiered feature selection strategy employing both MRMR and Relief algorithms was implemented to refine feature sets, followed by the development of an ensemble decision tree-based classifier. The model underwent rigorous multi-center validation and exhibited diagnostic performance on par with that of three experienced clinicians, underscoring its potential utility in clinical decision-making.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe model incorporating perinodular vascular features significantly outperformed the non-vascular feature model. Furthermore, the combined clinical-vascular-radiomic model demonstrated substantially improved performance over the clinical-vascular model, achieving AUCs of 0.8704 (CI: [0.6417,0.9676]) validation set, 0.8225 on independent test set (CI: [0.7298,0.9168]), and 0.7937 (CI: [0.4234,1]) on external test set. The PVC feature was consistently identified as one of the most important feature among all features in both feature selection and SHAP interpretability analysis.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e\u003cp\u003eIntegration of vascular characteristics markedly improves diagnostic performance and model generalizability. The consistent importance of PVC highlights its clinical value, and the model shows promising potential to assist in decision-making and reduce unnecessary invasive procedures.\u003c/p\u003e","manuscriptTitle":"Improving Risk Stratification of Pulmonary Nodules: An Integrated Perinodular Vascular and Radiomic Model for Clinical Decision Support","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-16 19:55:09","doi":"10.21203/rs.3.rs-8175303/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-16T07:34:51+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-13T19:09:26+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-07T18:53:14+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-06T18:24:34+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"331186190242337674750604791927921440273","date":"2026-02-06T13:42:11+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"242355881290443213248523810005147249461","date":"2026-02-04T06:48:00+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"63419649891422074634087694837857816544","date":"2026-01-19T16:27:03+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-10T02:21:07+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-11-27T07:41:48+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-26T10:54:03+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-11-26T10:53:49+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-11-21T16:07:01+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9be0e1e1-b392-4c42-8b5f-399093f452c5","owner":[],"postedDate":"December 16th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-04-13T15:59:58+00:00","versionOfRecord":{"articleIdentity":"rs-8175303","link":"https://doi.org/10.1186/s12880-026-02299-y","journal":{"identity":"bmc-medical-imaging","isVorOnly":false,"title":"BMC Medical Imaging"},"publishedOn":"2026-04-10 15:57:19","publishedOnDateReadable":"April 10th, 2026"},"versionCreatedAt":"2025-12-16 19:55:09","video":"","vorDoi":"10.1186/s12880-026-02299-y","vorDoiUrl":"https://doi.org/10.1186/s12880-026-02299-y","workflowStages":[]},"version":"v1","identity":"rs-8175303","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8175303","identity":"rs-8175303","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}