The diagnostic accuracy of artificial intelligence models in detecting lymph node metastases in lung cancer using endobronchial ultrasound (EBUS) images: A bivariate meta-analysis.

PURPOSE: This study systematically evaluated AI models for detecting lymph node metastases in lung cancer using EBUS images and assessed their role in thoracic oncology.

MATERIALS AND

METHODS

A systematic search following PRISMA-DTA guidelines was conducted in PubMed, Embase, Scopus, and Web of Science. Studies using AI models with cytologic or histologic analysis as the reference standard were included (PROSPERO: CRD42025635581). A bivariate random-effects model pooled sensitivity, specificity, diagnostic odds ratio (DOR), and area under the curve (AUC). AI models were categorized as CNN-A (Convolutional Neural Networks), BASM (Biomedical Application-Specific Models), AFP (Automated Frameworks and Platforms), and G-DDN (Generic Deep Neural Networks).

RESULTS

Twenty-two studies were included. The pooled sensitivity was 0.87 (95 % CI: 0.68-0.95), specificity 0.90 (95 % CI: 0.83-0.94), AUC 0.94 (95 % CI: 0.92-0.96), and DOR 56 (95 % CI: 17-182). CNN-A showed the highest accuracy, with an AUC of 0.970 and a DOR of 182, while AFP had the lowest sensitivity (0.058) and DOR (5.125), suggesting limited clinical applicability. Likelihood ratios were LR+ 8.39 (95 % CI: 4.93-14.28) and LR- 0.15 (95 % CI: 0.06-0.39), corresponding to post-test probabilities of 74 % for positive and 5 % for negative results. Subgroup analyses highlighted performance variations, emphasizing the need for refinement and validation in diverse settings.

CONCLUSION

AI models demonstrate high diagnostic accuracy in detecting lymph node metastases in lung cancer using EBUS images, reinforcing their potential in clinical decision-making. Future studies should refine accuracy metrics and further evaluate CNN-A across disease contexts.