Diagnostic performance of artificial intelligence versus conventional imaging for differentiating G2/G3 from G1 pancreatic neuroendocrine tumors: a systematic review and meta-analysis

BACKGROUND: Accurate preoperative grading of pancreatic neuroendocrine tumors (PanNETs), specifically differentiating G2/G3 from G1, is pivotal for treatment planning but challenging with conventional biopsy. This study aims to evaluate the diagnostic performance of imaging modalities, particularly comparing Machine Learning/Deep Learning (ML/DL) algorithms against conventional expert interpretation. METHODS: We conducted a systematic review and meta-analysis of studies from PubMed, Embase, Web of Science, and Cochrane Library up to December 2025. Studies utilizing CT, MRI, or endoscopic ultrasound (EUS) for PanNET grading were included. A bivariate random-effects model was used to calculate pooled sensitivity, specificity, and area under the curve (AUC). Subgroup analyses were performed to investigate heterogeneity and the impact of validation strategies. RESULTS: Seventeen studies comprising 928 patients were included. The overall pooled sensitivity and specificity were 0.77 (95% CI: 0.71–0.83) and 0.83 (95% CI: 0.77–0.87), respectively, with an AUC of 0.87. Notably, ML/DL models demonstrated significantly higher sensitivity than conventional imaging (0.84 vs. 0.71, p < 0.01) but lower specificity (0.78 vs. 0.87, p < 0.01). Multicenter studies showed a trend toward higher diagnostic metrics compared to single-center cohorts. Interestingly, the performance gap between external and internal validation narrowed when restricted to AI subgroups, suggesting the robustness of modern algorithms. CONCLUSIONS: Imaging-based analysis offers high diagnostic accuracy for PanNET grading. A theoretical sequential diagnostic strategy is suggested: utilizing AI’s high sensitivity for initial screening, followed by expert radiological review to ensure specificity.