Deep learning-assisted fine assessment of pulmonary nodule images

With the increasing adoption of low-dose computed tomography (LDCT) screening and the widespread use of chest CT in health examinations, chronic disease management, and multidisciplinary care, the detection rate of pulmonary nodules has risen substantially. Accordingly, the focus of pulmonary nodule management has shifted from simple lesion detection to refined evaluation, risk stratification, and longitudinal follow-up. Traditional radiologic assessment mainly relies on nodule diameter, density, margin characteristics, and interval growth. Although these approaches remain clinically valuable, they are limited by interobserver variability, suboptimal reproducibility, and difficulty in longitudinal comparison, especially in small nodules, juxta-vascular nodules, pleural-based nodules, subsolid nodules, and multiple nodules. Deep learning can automatically extract multi-scale imaging representations from two-dimensional, three-dimensional, and longitudinal CT data. In recent years, it has been widely applied to pulmonary nodule detection, precise segmentation, phenotypic characterization, malignancy risk prediction, dynamic follow-up, and progression forecasting. This review summarizes the clinical basis of refined pulmonary nodule imaging assessment and the current management framework, and systematically discusses recent advances in deep learning for nodule detection and segmentation, radiologic phenotype characterization, malignancy risk stratification, longitudinal follow-up, and clinical translation. In addition, based on the Fleischner Society guidelines, ACR Lung-RADS, BTS guideline, and recent consensus statements on subsolid nodules, this review analyzes the major barriers to real-world implementation, including insufficient external validation, heterogeneity of reference standards, limited interpretability, poor probability calibration, and incomplete workflow integration.

Key Words: pulmonary nodule; deep learning; low-dose computed tomography; ground-glass nodule; refined imaging assessment; risk stratification