JMIR Publications

ABSTRACT

Background:

Accurate preoperative assessment of pelvic lymph node metastasis (PLNM) is essential for treatment planning and prognostic stratification in prostate cancer (PCa). Although extended pelvic lymph node dissection (ePLND) remains the reference standard for nodal staging, it is invasive and associated with procedure-related morbidity. Imaging-based artificial intelligence (AI) has emerged as a promising noninvasive adjunct for preoperative nodal risk stratification.

Objective:

This study aimed to evaluate the diagnostic performance and clinical utility of imaging-based AI models for the preoperative prediction of PLNM in PCa.

Methods:

A systematic review and diagnostic test accuracy meta-analysis was conducted in accordance with the PRISMA-DTA statement. PubMed, Embase, Web of Science, and the Cochrane Library were searched from inception to January 2026. Study quality was assessed using QUADAS-2. For internal validation cohorts, pooled sensitivity and specificity were estimated using a bivariate random-effects model within the hierarchical summary receiver operating characteristic framework. Because only 3 external validation cohorts were available, sensitivity and specificity for external validation were synthesized separately using random-effects models. Clinical utility was evaluated using Fagan nomograms, and exploratory subgroup analyses were performed to investigate potential sources of heterogeneity.

Results:

Fourteen studies were included. In the internal validation cohorts, pooled sensitivity and specificity were 0.89 (95% CI 0.77-0.95) and 0.74 (95% CI 0.60-0.84), respectively, with a summary area under the curve of 0.89 (95% CI 0.86-0.91). In the external validation cohorts (n=3), pooled sensitivity and specificity were 0.79 (95% CI 0.42-0.95) and 0.72 (95% CI 0.61-0.81), respectively. At a pretest probability of 20%, a positive AI result increased the posttest probability to 46% in the internal validation cohorts and 41% in the external validation cohorts, whereas a negative result reduced it to 4% and 7%, respectively. Exploratory subgroup analyses suggested broadly comparable diagnostic performance between MRI-based and PET/CT-based models. Hybrid models showed stable pooled performance, whereas radiomics-only studies were too sparse for robust pooling. No significant publication bias was detected in the internal validation cohorts.

Conclusions:

Imaging-based AI models demonstrate promising diagnostic performance as noninvasive adjunctive tools for the preoperative prediction of PLNM in PCa. However, substantial heterogeneity, sparse external validation, and modest performance attenuation in external cohorts indicate that further standardized prospective multicenter studies are needed before routine clinical implementation. Clinical Trial: PROSPERO CRD420251274214