Advances in artificial intelligence for benign-malignant differentiation and risk-stratified management of pulmonary nodules
- VernacularTitle:人工智能驱动肺结节良恶性鉴别与风险分层管理
- Author:
Xuhua HUANG
1
;
Yiqing WANG
1
;
Xudong LÜ
2
;
Jianmin WU
3
,
4
;
Ping WANG
5
;
Wang LÜ
1
;
Jian HU
1
,
6
;
Luming WANG
1
Author Information
1. Department of Thoracic Surgery, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, 310003, P. R. China
2. College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, 310027, P. R. China
3. Department of Chemistry, Zhejiang University, Hangzhou, 310027, P. R. China
4. Future Medical Equipment Laboratory, Innovation Center of Yangtze River Delta, Zhejiang University, Hangzhou, 310027, P. R. China
5. Hangzhou Smart Intelligent Technology Co., Ltd., Hangzhou, 311100, P. R. China
6. Key Laboratory of Clinical Evaluation Technology for Medical Devices of Zhejiang Province, Hangzhou, 310003, P. R. China
- Publication Type:Journal Article
- Keywords:
Artificial intelligence;
pulmonary nodules;
benign-malignant differentiation;
risk stratification;
radiomics;
deep learning
- From:
Chinese Journal of Clinical Thoracic and Cardiovascular Surgery
2026;33(10):1559-1572
- CountryChina
- Language:Chinese
-
Abstract:
With the increasing use of low-dose computed tomography (LDCT) in lung cancer screening and health examinations, the clinical goal of pulmonary nodule management has shifted from detecting nodules to identifying lesions that warrant intervention while avoiding overdiagnosis and overtreatment of low-risk disease. Artificial intelligence (AI) has consequently evolved from computer-aided detection and segmentation to benign-malignant differentiation, longitudinal growth assessment, invasiveness prediction, risk stratification, and closed-loop workflow support. This review outlines the developmental trajectory of AI for pulmonary nodules and summarizes advances in detection and segmentation, radiomics, end-to-end deep learning, longitudinal modeling, and multimodal integration across incidentally detected, screening-detected, subsolid, and multiple-nodule scenarios. Particular attention is given to real-world failure modes and their causes, including data and label bias, scanner- and protocol-related distribution shift, spectrum bias and overfitting in subsolid nodules, missed atypical lesions, automation bias, and management of discordance between AI outputs and guideline-based recommendations. Current evidence suggests that AI may improve detection efficiency and risk reclassification in selected settings; however, evidence regarding cross-population calibration, patient-relevant outcomes, cost-effectiveness, and post-deployment monitoring remains limited. AI should therefore be positioned as a human-in-the-loop adjunct within guideline-governed pathways, supported by scenario-specific validation, transparent reporting, interpretability and accountability, continuous performance auditing, and multidisciplinary decision-making.
