APEX-NET: automated pancreatic evaluation network using early non-contrast CT

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Acute pancreatitis (AP) diagnosis and severity assessment often rely on contrast-enhanced CT (CECT), which involves a delay.
• Early non-contrast CT (NCCT) is the initial scan but lacks detailed information for accurate severity stratification.
• There is a need for a method to predict AP severity from early NCCT scans to optimize treatment timing.

Purpose of the Study:

• To develop and validate APEX-NET, a deep learning model for early diagnosis and severity stratification of AP using NCCT.
• To leverage CECT feature learning to derive simulated CECT features from NCCT scans.
• To assess APEX-NET's performance against existing methods and radiologists.

Main Methods:

• A five-center study involving 3383 patients (AP and Non-AP).
• APEX-NET was trained for pancreas segmentation, AP diagnosis, and severity prediction (mild, moderately severe, severe).
• A feature mapping module derived simulated CECT features from NCCT using paired NCCT-CECT data; model validated across internal and external cohorts.

Main Results:

• APEX-NET achieved high AUCs for AP diagnosis (0.949-0.981) and severity prediction (0.873-0.872).
• The model demonstrated significant outperformance over NCCT-only models for severity prediction.
• APEX-NET showed comparable performance to senior radiologists and superior performance to junior radiologists in a reader study.

Conclusions:

• APEX-NET enables accurate, NCCT-based diagnosis and early severity stratification of AP.
• The model shows strong potential for clinical integration, overcoming CECT delays.
• This deep learning framework accelerates AP assessment and can be integrated into clinical workflows.