Background: Infected pancreatic necrosis significantly increases morbidity and mortality in acute necrotizing pancreatitis (ANP). The only indicator of infection on CT scans is gas, which demonstrates poor sensitivity. This study aimed to develop and validate a machine learning model based on CT radiomic features to predict infected pancreatic necrosis and compare its performance with conventional markers.
Methods: We retrospectively analyzed data from 75 patients with moderately severe and severe ANP who underwent CT imaging during the second or third week of illness prior to any intervention. The final diagnosis of infected pancreatic necrosis was confirmed by microbiological culture of pancreatic fluid obtained during the first endoscopic or percutaneous intervention. Radiomic features were extracted from pancreatic collections on baseline CT scans. The cohort was split into training (n=60) and test (n=15) sets. A machine learning algorithm was trained to predict infected necrosis, and its performance was compared with conventional predictors including gas on CT, procalcitonin (PCT), and C-reactive protein (CRP).
Results: The study cohort (mean age 38.5 years, 80% male) included 43 patients (57.3%) with culture-confirmed infected pancreatic necrosis. Most patients (65.3%) had severe disease with 60% experiencing organ failure. Mean collection size was 7.8±3.2 cm. Gas was visible on CT in only 18 patients (24%), demonstrating a sensitivity of 34.9% and specificity of 87.5% (AUC: 0.61) for predicting infection. PCT achieved an AUC of 0.68 with optimal cutoff at 1.5 ng/ml (sensitivity 62.8%, specificity 71.9%), while CRP showed an AUC of 0.57 with optimal cutoff at 250 mg/l (sensitivity 65.1%, specificity 46.9%). The radiomics-based machine learning model demonstrated [AUC of 0.71 (95% CI: 0.61-0.76)] for predicting infected necrosis, significantly outperforming all conventional markers. When combining radiomic features with PCT, diagnostic accuracy improved further to [AUC: 0.75 (95% CI: 0.66-0.79)].
Conclusion: CT-based radiomic features analyzed through machine learning demonstrate superior performance for predicting infected pancreatic necrosis compared to conventional markers including gas on CT, PCT, and CRP. This approach has the potential to facilitate earlier diagnosis and timely intervention in patients with ANP, potentially improving clinical outcomes. Further prospective validation in larger cohorts and external test set is warranted.