Objectives To develop and validate an automated radiomics-based model to objectively assess the malignancy risk of pulmonary nodules, overcoming the limitations of manual CT interpretation. Materials and methods This prospective, multicenter diagnostic accuracy study enrolled 1895 patients with 1909 pulmonary nodules (1181 malignant; 728 benign) from 27 centers between 2017 and 2023. Clinical data and chest CT images were collected, and 25 radiological and 2153 radiomics features were extracted after 3D U-net–based segmentation. Three predictive models were developed: clinical–radiological (“Human Reading”), radiomics-only (“Radiomics”), and a combined model. Nodules were divided into training (n = 830), internal validation (n = 214), and external validation (n = 865) sets. The primary endpoint was diagnostic accuracy, assessed by AUC. Results Participants included 888 men and 1007 women (mean age, 54.8 ± 11 years). In internal validation, the human reading and radiomics models achieved similar performance (AUC 0.88 [95% CI: 0.82–0.94] vs 0.88 [0.83–0.93]; p = 0.87). External validation confirmed comparable results (AUC 0.86 [0.83–0.88] vs 0.85 [0.82–0.87]; p = 0.56). The combined model outperformed both (AUC gain + 2.4% [vs radiomics], p < 0.001; + 1.7% [vs human reading], p = 0.0025). Conclusion Integrating radiomics with clinical–radiological features enhances pulmonary nodule malignancy prediction, offering an effective and scalable tool for lung cancer risk assessment, particularly where radiological expertise is limited. Clinical trial registration NCT03181490, NCT03651986. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Can an automated radiomics-based model accurately predict the malignancy risk of pulmonary nodules, reducing the subjectivity and workload of manual CT interpretation? <Emphasis Type="BoldItalic">Findings</Emphasis> In a multicenter cohort of 1895 patients, a combined radiomics–clinical model achieved the highest diagnostic accuracy (AUC 0.87–0.90), outperforming human reading alone. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> Integrating radiomics with clinical–radiological features enables objective and scalable lung cancer risk assessment, potentially improving early detection and diagnostic consistency across diverse clinical settings. Graphical Abstract

Comparison of clinical–radiological and radiomics features for predicting pulmonary nodule malignancy in a multicenter study of mixed clinical and surveillance populations

期刊论文
发表日期：

2026年3月14日

查看全文

Fanrui Zeng,
Bo Wang,
Minhua Peng,
Jinsheng Tao,
Xixiang Tu,
Hongbo Lu,
Xiangcheng Qiu,
Yang Yang,
Wenyang Wang,
Meng Zeng,
Jian-bing Fan,
Qingsi Zeng,
Qin Liu

摘要

Objectives

To develop and validate an automated radiomics-based model to objectively assess the malignancy risk of pulmonary nodules, overcoming the limitations of manual CT interpretation.

Materials and methods

This prospective, multicenter diagnostic accuracy study enrolled 1895 patients with 1909 pulmonary nodules (1181 malignant; 728 benign) from 27 centers between 2017 and 2023. Clinical data and chest CT images were collected, and 25 radiological and 2153 radiomics features were extracted after 3D U-net–based segmentation. Three predictive models were developed: clinical–radiological (“Human Reading”), radiomics-only (“Radiomics”), and a combined model. Nodules were divided into training (n = 830), internal validation (n = 214), and external validation (n = 865) sets. The primary endpoint was diagnostic accuracy, assessed by AUC.

Results

Participants included 888 men and 1007 women (mean age, 54.8 ± 11 years). In internal validation, the human reading and radiomics models achieved similar performance (AUC 0.88 [95% CI: 0.82–0.94] vs 0.88 [0.83–0.93]; p = 0.87). External validation confirmed comparable results (AUC 0.86 [0.83–0.88] vs 0.85 [0.82–0.87]; p = 0.56). The combined model outperformed both (AUC gain + 2.4% [vs radiomics], p < 0.001; + 1.7% [vs human reading], p = 0.0025).

Conclusion

Integrating radiomics with clinical–radiological features enhances pulmonary nodule malignancy prediction, offering an effective and scalable tool for lung cancer risk assessment, particularly where radiological expertise is limited.

Clinical trial registration

NCT03181490, NCT03651986.

Key Points

Question Can an automated radiomics-based model accurately predict the malignancy risk of pulmonary nodules, reducing the subjectivity and workload of manual CT interpretation?

Findings In a multicenter cohort of 1895 patients, a combined radiomics–clinical model achieved the highest diagnostic accuracy (AUC 0.87–0.90), outperforming human reading alone.

Clinical relevance Integrating radiomics with clinical–radiological features enables objective and scalable lung cancer risk assessment, potentially improving early detection and diagnostic consistency across diverse clinical settings.

Graphical Abstract