Uncertainty-Aware Explainable AI for Pancreatic Cysts: Identifying Deep Learning Vulnerabilities and Ensuring Safe...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging
• Oncology

Background:

• Growing interest in AI for pancreatic cyst management.
• Lack of systematic investigation into how lesion characteristics influence AI model behavior for IPMN malignancy risk stratification.
• Need for comprehensive explainability analysis in AI-driven IPMN assessment.

Purpose of the Study:

• To systematically investigate how cyst characteristics (type, size, location) affect deep learning model performance in IPMN malignancy prediction.
• To integrate explainable AI (XAI) with uncertainty quantification for evaluating AI model behavior.
• To assess the impact of lesion characteristics on AI model interpretability and performance in IPMN malignancy risk stratification.

Main Methods:

• Retrospective multi-center study of 170 IPMNs using a radiomics-deep learning fusion model.
• Stratification of cases by dysplasia grade, IPMN type, size, and location.
• Model interpretability assessed using SHAP, LIME, Grad-CAM, and uncertainty quantification.

Main Results:

• Overall accuracy of 67.1%; model uncertainty was lower for correct vs. incorrect predictions (0.72 vs. 0.78, p < 0.001).
• Subgroup analysis revealed higher accuracy for low-risk (81.3%) vs. high-risk IPMNs (42.9%).
• Performance varied by cyst type (BD inversely correlated with size) and location (body/whole-pancreas > tail).

Conclusions:

• AI models exhibit significant performance drops with complex, high-risk IPMNs.
• Uncertainty quantification effectively flags unreliable AI predictions.
• A selective-prediction framework using AI and uncertainty quantification can guide clinical decisions for IPMN management.