An endoscopic ultrasound-based interpretable deep learning model and nomogram for distinguishing pancreatic neuroendocrine tumors from pancreatic cancer

Abstract

Objectives To retrospectively validate and develop an interpretable deep learning model and nomogram using EUS images to predict pancreatic neuroendocrine tumors (pNETs). Methods After pathological confirmation, a retrospective analysis of 266 patients (115 with pNETs and 151 with pancreatic cancer) was conducted. Patients were randomly divided into training and test groups (7:3 ratio). The least absolute shrinkage and selection operator algorithm reduced DL feature dimensions from pre-standardized EUS images. Nonzero features developed eight predictive DL models using different machine learning algorithms. The best model established a clinical signature for a nomogram. Grad-CAM and SHAP were used to interpret and visualize model outputs. Results Out of 2048 DL features, 27 with nonzero coefficients were retained. The SVM DL model achieved AUCs of 0.948 (training) and 0.795 (test). A nomogram combining DL and clinical signatures was developed, and calibration curves, DCA plots, and CICs confirmed high accuracy. Grad-CAM and SHAP enhanced model interpretability, benefiting clinical decision-making. Conclusions The novel interpretable DL model and nomogram, validated with EUS images and machine learning, show promise for enhancing EUS's ability to predict pNETs from pancreatic cancer, providing valuable insights for future research and application.