Deep learning-based prediction of acute pancreatitis severity from abdominal CT with multicenter external validation

Area of Science:

• Radiology
• Artificial Intelligence
• Gastroenterology

Background:

• Acute pancreatitis (AP) is a common condition with increasing incidence.
• Severe AP (SAP) has high mortality, necessitating early and accurate severity prediction.
• Current models like BISAP and mCTSI have limitations in accuracy and data availability at admission.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for predicting AP severity.
• The model utilizes abdominal contrast-enhanced CT scans obtained within 24 hours of patient admission.

Main Methods:

• A large dataset of 10,130 studies from 8,335 patients was used, including multi-site U.S. health system data and public datasets.
• The DL model underwent a two-stage training process: self-supervised pretraining on unlabeled data and fine-tuning on labeled data.
• Performance was evaluated against mCTSI and BISAP on internal and external test sets.

Main Results:

• The DL model demonstrated high performance, with AUROCs of 0.888 for SAP and 0.888 for MAP on the internal test set, outperforming mCTSI.
• External validation confirmed robust performance with AUROCs of 0.887 for SAP and 0.858 for MAP, surpassing both mCTSI and BISAP.
• The model showed effectiveness in retrospective triage analysis, identifying a significant proportion of patients with varying AP severity.

Conclusions:

• The developed DL model offers comparable or superior performance to existing prognostic tools for AP severity.
• The model exhibits strong external validation, indicating its generalizability.
• AI-assisted CT analysis holds promise for early, automated risk stratification in acute pancreatitis.