Abstract
<title>Abstract</title> <p> Background Preoperative identification of lymph node metastasis (LNM) is pivotal for tailoring surgical strategies in prostate cancer. Current radiologic assessment is limited by low sensitivity for micro-metastasis. Artificial intelligence (AI) including machine learning (ML) and deep learning (DL), has been increasingly applied, but their comparative diagnostic performance remains unclear. Methods This meta-analysis was performed in accordance with PRISMA-DTA guidelines. Databases were searched from inception to February 2026. Bivariate random-effects models were used to pool diagnostic metrics. Subgroup analyses and meta-regression were conducted. Results A total of 21 studies (15 ML, 6 DL) comprising 2,054 patients were included. The overall pooled sensitivity was 0.81 (95% CI: 0.73–0.87), specificity was 0.84 (95% CI: 0.78–0.89), positive likelihood ratio (PLR) was 5.14 (3.62–7.29), negative likelihood ratio (NLR) was 0.23 (0.15–0.33), diagnostic odds ratio (DOR) was 22.79 (12.84–40.46), and the (AUC) was 0.90 (95% CI: 0.87–0.92). ML and DL showed comparable overall accuracy (0.80 vs. 0.83). A striking algorithm–modality interaction was identified: in DL models, magnetic resonance imaging (MRI) was superior to computed tomography (CT) (Accuracy: 0.91 vs. 0.78, <italic>P</italic> = 0.004); in ML models, CT was superior to MRI (Accuracy: 0.85 vs. 0.73, <italic>P</italic> = 0.043). Subgroup meta-regression confirmed imaging modality as a primary source of heterogeneity within ML and DL groups. Deeks’ test showed no publication bias ( <italic>P</italic> = 0.65). Conclusion Both ML and DL models demonstrated excellent diagnostic performance in prostate cancer LNM. Clinical implementation should follow an algorithm–modality matching principle: DL is more suitable for MRI data, while ML is optimized for CT data. </p>