Lung cancer remains a significant global public health threat, with the latest epidemiological data from 2024 indicating that its mortality rate continues to rank first among malignant tumors (1). In China, the reported 5-year survival rate for lung cancer is 28.7, which remains relatively low (2, 3). However, the prognosis of lung cancer varies considerably depending on its stage, with the 5-year survival rate reaching up to 92% for stage I lung cancer (4). Nevertheless, the delayed diagnosis of pulmonary nodules (PNs) due to the absence of typical symptoms in their early stages often leads to patients being diagnosed at advanced stages, highlighting the urgent clinical need for the development of precise early screening techniques.

In the field of differentiating benign and malignant PNs, significant advancements have been made in multimodal diagnostic techniques. With the adoption of low-dose computed tomography (LDCT) as the primary screening modality, a 20% reduction in mortality has been achieved among high-risk populations. However, LDCT screening is prone to false-positive diagnoses or underdiagnosis, the baseline false-positive rate was reported as 24% in the National Lung Screening Trial (NLST) (5). Nevertheless, with advancements in imaging technology and follow-up strategies, the postoperative pathological benign rate has progressively declined from 14.5% to 6.2% (6). Radiologists traditionally identify PNs through manual interpretation of chest CT scans—a process that is time-consuming, labor-intensive, and susceptible to diagnostic variability.

In this context, the “uAI platform,” a PNs intelligent diagnostic system developed by Shanghai United Imaging Company, has overcome the limitations of traditional computer-aided diagnosis (CAD). This system is based on a supervised deep transfer learning (SDTL) framework trained on high-quality annotated CT images. It achieves a diagnostic AUC of 91.8% for pulmonary nodules ≤3 cm in size, improves model specificity by 12.3% (7). The uAI platform enables highly sensitive detection of PNs, providing quantitative analysis of multidimensional parameters including PNs size, volume, and composition. This system facilitates comprehensive and precise evaluation of lesion characteristics while offering intelligent prediction of PNs benign/malignant status.

Circulating tumor cells (CTCs) are tumor cells that detach from the primary or metastatic lesions and enter the bloodstream or lymphatic system. These cells have potential value in the diagnosis and monitoring of malignant tumor progression, particularly exhibiting unique advantages in the early diagnosis of lung adenocarcinoma. Studies have already utilized CTCs for the diagnosis of lung adenocarcinoma and lung cancer subtyping (8, 9). However, no research has described their association with the degree of lung cancer invasion. Additionally, the techniques for capturing and isolating CTCs are diverse, and the sensitivity of detection is often limited by the volume of blood samples, such as the use of EpCAM for positive or negative enrichment (10, 11). The GILUPI CellCollector, an in-vivo examination technology equipped with a fully automated operating system, enables the counting of CTCs and the detection of tumor-specific protein subtypes, significantly improving the efficiency and accuracy of detection (12).

This study innovatively integrates the PNs risk stratification of the uAI platform with quantitative detection of CTCs, constructing a PNs diagnostic model. It is the first to confirm the synergistic effect of the combined strategy of artificial intelligence and liquid biopsy on the stratification of malignant risk in PNs, providing a practical solution for advancing the early diagnosis and treatment of lung adenocarcinoma.

In recent years, the incidence and mortality rates of lung malignancies have ranked first globally. Studies indicate that the recurrence rate of early-stage lung cancer is only 10% five years after surgery, highlighting the importance of early detection and surgical intervention in the treatment of lung cancer (13). Early-stage lung cancer does not exhibit specific clinical manifestations, and its initial presentation may merely be an isolated pulmonary nodule. Based on the solid components within the nodule, isolated nodules can be categorized into solid nodules and subsolid nodules. Solid nodules refer to those where the nodule density is sufficient to obscure the small bronchi and blood vessels within, while subsolid nodules are characterized by unclear or indefinite boundaries, and the nodule density does not obscure the small bronchi and blood vessels passing through it (14, 15).

Studies have demonstrated that compared to non-small cell lung cancer (NSCLC) presenting as subsolid nodules on imaging, pulmonary lesions manifesting as pure solid nodules exhibit higher rates of lymph node metastasis, greater malignant potential, and poorer prognosis (16, 17). However, the radiological characteristics of pulmonary nodules (PNs) with diameters ≤1 cm are often inconspicuous, leading to interobserver variability in interpretation due to differences in clinical experience (18, 19). In recent years, numerous studies have explored the clinical application of AI-based radiological models for early lung cancer prediction, providing valuable assistance in characterizing PNs. Catelli et al. utilized high-resolution computed tomography (HRCT) features (e.g., spiculation, size, and density) to differentiate benign from malignant nodules, offering clinicians a clearer reference for determining the nature of PNs and formulating surgical strategies (20). Pan et al. proposed an innovative approach by integrating binary and ternary classification models with a pruning decision strategy to resolve classification conflicts, thereby improving the accuracy of predicting the invasiveness of lung adenocarcinoma and demonstrating its potential for CT-based risk stratification in lung cancer (21). Nevertheless, CT radiological characteristics alone cannot fully distinguish benign from malignant PNs. Previous studies have indicated that combining radiological features with additional and more sensitive molecular markers can enhance the diagnostic positivity rate for lung cancer (22–24). Therefore, in this study, we employed uAI platform and CTC assisted analysis to uniformly assess the malignancy risk of patients’ PNs.

Relevant studies have indicated that certain tumor markers in serum can also aid in the diagnosis of lung cancer. However, traditional tumor markers such as CEA, NSE, and CA199 demonstrate limited specificity and sensitivity in the early diagnosis of lung cancer (24–27). In recent years, the emergence of liquid biopsy techniques, represented by CTCs, has provided a novel approach for the early diagnosis of lung cancer. CTCs refer to tumor cells that detach from primary or metastatic lesions and enter the bloodstream. The number of CTCs in peripheral blood is extremely low, and specific CTC markers are crucial for improving CTC detection rates. Folate receptor positivity, as a specific marker for CTCs, exhibits high sensitivity and specificity and is highly expressed in tumor cells. The folate receptor is minimally expressed in the cells of the fallopian tube, renal tubules, alveolar walls, choroid, and uterus, and is not expressed in blood cells. Its expression in lung cancer cells exceeds 78%. Furthermore, the folate receptor can identify active CTCs, unaffected by the transition from epithelial to mesenchymal cells. In this study, the CellCollector (GILUPI CellCollector, GILUPI) was used for in-vivo detection of peripheral CTCs before surgery. CTC assessment analysis were performed using immunofluorescence staining. Additionally, CTCs captured by the CellCollector sampling probe were isolated for whole-genome amplification and quality assessment. In 2010, the American Joint Committee on Cancer (AJCC) Cancer Staging Manual proposed that CTCs could serve as a new indicator to assist pathologists in pathological staging. Clinically, CTCs have begun to be used to better develop treatment plans for lung cancer patients (28–30). Previous attempts have been made to combine CTC assessment with established screening methods, focusing on improving specificity by integrating low-dose CT screening programs with subsequent CTC evaluations. CTC enumeration assessments were conducted on patients with identified “ground-glass” lung nodules and healthy controls. CTCs were only present in the blood of some patients with nodules. Based on subsequent molecular analysis, these CTCs were found to have a “malignant tendency” (31). However, the study did not establish a model to predict the pathological nature of nodules and lacked a unified set of diagnosis codes. We retrospectively evaluated the diagnostic efficacy of combining liquid biopsy with artificial intelligence in 76 patients suspected of having lung cancer. The results demonstrated the superior predictive performance of this forecasting model.

This study has several limitations. Firstly, the pulmonary nodules detected in this study were exclusively focused on lung adenocarcinoma; future research should incorporate pathological data from other types of malignant nodules. Second, due to the retrospective nature of this study and the relatively small sample size collected, prospective validation of the hypotheses was unattainable. Further large-scale, multicenter studies are warranted for conclusive verification.

This study developed a pulmonary nodule prediction model integrating uAI platform and CTCs, demonstrating robust accuracy, stability, and clinical applicability through internal validation. In the future, we will verify it through large-sample, multi-center studies.