LAUSR

This study aimed to develop and validate a combined nomogram model based on deep learning (DL) contrast-enhanced ultrasound (CEUS) and clinical factors to preoperatively predict the aggressiveness of pancreatic neuroendocrine neoplasms (PNENs). In this retrospective study, consecutive patients with histologically proven PNENs underwent CEUS examination at the initial work-up between January 2010 and October 2020. Patients were randomly allocated to the training and test sets. Typical sonographic and enhanced images of PNENs were selected to fine-tune the SE-ResNeXt-50 network. A combined nomogram model was developed by incorporating the DL predictive probability with clinical factors using multivariate logistic regression analysis. The utility of the proposed model was evaluated using receiver operator characteristic, calibration, and decision curve analysis. A total of 104 patients were evaluated, including 80 (mean age ± standard deviation, 47 years ± 12; 56 males) in the training set and 24 (50 years ± 12; 14 males) in the test set. The DL model displayed effective image recognition with an AUC of 0.81 (95%CI: 0.62–1.00) in the test set. The combined nomogram model that incorporated independent clinical risk factors, such as tumor size, arterial enhancement level, and DL predictive probability, showed strong discrimination, with an AUC of 0.85 (95%CI: 0.69–1.00) in the test set with good calibration. Decision curve analysis verified the clinical usefulness of the combined nomogram. The combined nomogram model could serve as a preoperative, noninvasive, and precise evaluation tool to differentiate aggressive and non-aggressive PNENs. • Tumor size (odds ratio [OR], 1.58; p = 0.02), arterial enhancement level (OR, 0.04; p = 0.008), and deep learning predictive probability (OR, 288.46; p < 0.001) independently predicted aggressiveness of pancreatic neuroendocrine neoplasms preoperatively. • The combined model predicted aggressiveness better than the clinical model (AUC: 0.97 vs. 0.87, p = 0.009), achieving AUC values of 0.97 and 0.85 in the training set and the test set, respectively.