S115 A Systematic Review and Meta-Analysis of Artificial Intelligence vs Clinicians for Pancreatic Cancer Diagnosis Using Computerized Tomography

Introduction: Pancreatic cancer, with its rising incidence and mortality imposes a profound clinical and economic burden. In the United States, direct healthcare costs due to pancreatic cancer exceed $2.5 billion. Pancreatic cancer is often diagnosed at an advanced stage due to non-specific early symptoms, limiting curative options. Early detection dramatically improves survival, but fewer than 10% of cases are identified before metastasis. Artificial intelligence (AI) has emerged as a transformative tool in medical diagnostics, with advanced deep learning neural networks now capable of detecting subtle, early-stage pancreatic malignancies beyond human capabilities. Methods: We performed a systematic review and meta-analysis comparing AI algorithms with clinicians for pancreatic cancer diagnosis using computed tomography imaging. Databases including Ovid Medline, CENTRAL, Embase, Scopus, Web of Science, and IEEE Xplore were searched from inception to present. Studies directly comparing AI and clinician diagnostic performance were included. Three reviewers independently screened, extracted data, and assessed bias using QUADAS-2. Random-effects meta-analyses and HSROC models were conducted in R, following PRISMA-DTA guidelines. Results: AI demonstrated superior diagnostic accuracy over clinicians in detecting pancreatic lesions. Pooled sensitivity and specificity for AI reached 95% and 96%, respectively, with lower inter-study heterogeneity compared to clinicians (sensitivity 87%, specificity 88%). In distinguishing malignant from benign pancreatic tumors, AI outperformed clinicians in sensitivity (93% vs 78%), specificity (75% vs 70%), positive predictive value (86% vs 80%), negative predictive value (87% vs 66%), and overall accuracy (86% vs 74%). HSROC analysis confirmed AI’s diagnostic superiority (AUC 0.874 vs 0.799), highlighting its potential to reduce diagnostic errors and minimize inter-observer variability. Conclusion: Our study affirms that AI-driven diagnostics consistently outperform clinicians in pancreatic cancer detection, echoing findings from recent experimental studies. AI integration reduces diagnostic variability, expedites workflows, and promises a paradigm shift in precision oncology. However, ethical considerations including transparency, bias mitigation, and data privacy remain critical for responsible implementation. Our findings provide robust, evidence-based guidance for clinicians and underscore the transformative potential of AI in pancreatic cancer care.