Automated CT Volumetry of Peripancreatic Collections for Risk Stratification in Acute Pancreatitis: A Multicenter Retrospective Study.
TL;DR
This deep learning framework enables rapid, fully automated PPC volumetry on CE-CT and provides clinically meaningful risk stratification for organ failure and infection in patients with acute pancreatitis and early peripancreatic collections.
Abstract
Rationale
AND
Objectives
Extrapancreatic necrosis volume is an established prognostic marker in acute necrotizing pancreatitis, yet early diagnosis remains challenging and manual segmentation is labor-intensive. We developed and validated an automated CT volumetry tool for peripancreatic collections (PPCs) on contrast-enhanced CT (CE-CT) and evaluated its utility for early risk stratification.
Methods
This retrospective study initially screened 520 patients with acute pancreatitis and ultimately included 394. Patients from the primary center (n = 303) were temporally divided into a development cohort (n = 198) for model construction and an internal test cohort (n = 105). An additional 91 patients from two external hospitals formed the external test cohort. Using manual segmentations verified by two radiologists as the reference standard, model performance was assessed via Dice coefficients and Pearson correlation. Multivariable logistic regression and receiver operating characteristic (ROC) analyses evaluated the association between automated PPC volume and the primary outcome (organ failure) as well as the exploratory secondary outcome (infection).
Results
The development, internal test, and external test cohorts included 198, 105, and 91 patients, respectively (mean age, 46 years; range, 18-90). In the external test cohort, automated segmentation showed excellent agreement with manual assessment [Dice coefficient, 0.89 (95% CI: 0.87-0.90); Pearson r = 0.99 (95% CI: 0.97-1.00; P < 0.001)]. Automated peripancreatic collection (PPC) volume (per 100 mL) was independently associated with organ failure and infection in both test cohorts (all P ≤ 0.006). In the external test cohort, automated PPC volume demonstrated strong performance for risk stratification of organ failure (AUC: 0.82 vs. modified CT severity index [mCTSI]: 0.73) and infection (AUC: 0.79 vs. mCTSI: 0.72). It also significantly outperformed mCTSI in the internal test cohort for both outcomes. The framework was substantially faster than manual segmentation (median, 25.2 s vs. 12.4 min per patient; P < 0.001) and surpassed radiologists' subjective binary classification (low vs. high PPC burden).
Conclusion
This deep learning framework enables rapid, fully automated PPC volumetry on CE-CT and provides clinically meaningful risk stratification for organ failure and infection in patients with acute pancreatitis and early peripancreatic collections.