Prediction of pancreatic neuroendocrine tumor grading using an artificial intelligence–based video analysis model (GradAINet) applied to contrast-enhanced EUS videos

Background and Objectives: Pancreatic neuroendocrine neoplasms (PNENs) are rare tumors with heterogeneous outcomes. Tumor grading (G), based on mitotic count and Ki-67 index, is the main prognostic factor guiding treatment. EUS–guided fine-needle aspiration/biopsy is the current standard but shows a misgrading rate up to 25%. We evaluated an artificial intelligence–based video analysis model to predict PNEN grading from contrast-enhanced EUS (CE-EUS) recordings. Methods: This retrospective study was conducted at Istituti di Ricovero e Cura a Carattere Scientifico San Raffaele Hospital, Milan, a European Neuroendocrine Tumor Society Center of Excellence. Patients were eligible if CE-EUS videos ≥1 minute (arterial and venous phases) and cyto-histological confirmation of PNEN were available. Exclusion criteria included mixed neuroendocrine–non-neuroendocrine neoplasms, missing Ki-67 grading, or poor video quality. CE-EUS videos were processed with a deep-learning video transformer model (GradAINet). The dataset was split into training (70%), validation (10%), and testing (20%) cohorts. Diagnostic performance was evaluated using sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and F1-score. Results: Between 2022 and 2024, 115 patients were included (49 female, 42.6%): 70 had G1, and 45 had G2–G3 tumors. Overall, 253,751 video frames were analyzed. GradAINet achieved a sensitivity of 0.817 (95% confidence interval [CI]: 0.556–1.000), specificity 0.806 (95% CI: 0.588–1.000), positive predictive value 0.759 (95% CI: 0.500–1.000), negative predictive value 0.856 (95% CI: 0.667–1.000), and accuracy 0.811 (95% CI: 0.654–0.962). Conclusion: This artificial intelligence–driven CE-EUS video model shows high accuracy for PNEN grading and potentially complements EUS-guided fine-needle aspiration/biopsy. As the first video-based rather than static-image model, it represents a methodological advance. Multicenter validation on larger cohorts is needed before clinical implementation.