Diagnostic performance of artificial intelligence-based imaging and liquid biopsy combined with clinical evaluation for distinguishing benign and malignant pulmonary nodules: a diagnostic network meta-analysis.

BACKGROUND: The incidence of pulmonary nodules is rising. In the United States alone, approximately 1.57 million cases are detected annually via low-dose CT screening. This growing number has created an urgent need for accurately differentiating malignant lesions. The current standard methods rely on radiologist interpretation. However, these methods show variable accuracy, with a sensitivity range of 62%-79%. Invasive biopsies, which are often used for further diagnosis, carry complication risks ranging from 15% to 28%. Emerging technologies, such as artificial intelligence (AI) imaging analysis and liquid biopsy platforms, show promising capabilities in pulmonary nodule characterization. Nevertheless, there is a lack of a rigorous methodology for their comparative evaluation.

METHODS: We systematically evaluated 20 comparative studies that were published between 2015 and 2024. These studies assessed AI-based imaging methods, which included deep learning architectures and radiomics analysis, as well as liquid biopsy methods, such as ctDNA mutation profiling and methylation-based assays, for the characterization of pulmonary nodules. Bayesian hierarchical modeling was employed to account for the interdependencies among different tests. Evidence certainty was assessed via the GRADE framework.

RESULTS: The CT-Deep approach, when integrated with clinical information, demonstrated the highest sensitivity, with a value of 97.1% and a 90.8%-99.6% credible interval. However, it showed relatively lower specificity, with a value of 67.4% and a 74.4%-92.9% credible interval. Stand-alone liquid biopsy exhibited more balanced operating characteristics, with a sensitivity of 63.1% and a specificity of 82.8%. The hybrid clinical-liquid biopsy strategy showed optimal performance in the summary receiver operating characteristic analysis, with an area under the curve of 0.903. Substantial heterogeneity was observed across studies, with an I2 > 50%. Additionally, there were wide confidence intervals for some diagnostic odds ratio estimates, which means that the comparative performance should be interpreted with caution.

CONCLUSIONS: AI-enhanced imaging techniques are particularly valuable for high-sensitivity screening applications, where the risk of false negatives is the greatest. On the other hand, liquid biopsy approaches or their combination with clinical assessment may be more suitable for situations that require high specificity. The findings highlight the need for standardized validation protocols and prospective evaluation of combined modality strategies to address the current limitations in pulmonary nodule characterization.