Acute pancreatitis is a common gastrointestinal disease with rising global incidence. Most cases are mild, but severe acute pancreatitis carries high mortality and can require early intensive management. Timely severity prediction can help distinguish patients likely to need closer monitoring from those who may follow a milder course. Current prognostic tools, including the Bedside Index of Severity in Acute Pancreatitis and the Modified CT Severity Index, have recognised limitations. Some depend on clinical information that may not be available at admission, while imaging-based scoring can vary between readers.
A 2026 analysis published in Radiology Advances assessed a deep learning model for predicting acute pancreatitis severity from admission contrast-enhanced CT scans. The model uses routinely acquired CT imaging within the first 24 hours of admission to support automated risk assessment before complications become fully apparent.
Training Across Multiple Datasets
The model estimates the probability of mild acute pancreatitis and severe acute pancreatitis from abdominal contrast-enhanced CT scans. Development used data from a multi-hospital academic system in the United States and several public abdominal CT datasets. Training took place in two stages. First, self-supervised learning used a large collection of unlabelled CT examinations. Then supervised fine-tuning used a smaller labelled set, with additional pseudo-labels extracted from radiology reports. The pseudo-labelled data supported fine-tuning only and did not enter validation or testing. During training, an automated segmentation tool helped identify pancreas-containing slices. At inference, segmentation was not required, allowing the model to operate directly on CT scans.
Must Read: Accelerating Pancreatic MRI with DL Reconstruction
The internal evaluation pathway included adults whose radiology reports were consistent with acute pancreatitis. Diagnostic confirmation and severity assignment followed the Revised Atlanta Classification. The internal test set included 100 patients. External validation used 518 CT studies from the GOULASH trial, a multicentre randomised controlled trial conducted at three tertiary medical centres in Hungary. This design tested whether the model could retain performance beyond the health system used for development. The external dataset also provided comparison with both the Modified CT Severity Index and the Bedside Index of Severity in Acute Pancreatitis.
Results Compared with Existing Scores
Results showed stronger discrimination with AI than with the Modified CT Severity Index in the internal test set. For both severe and mild acute pancreatitis, the AI model reached an area under the receiver operating characteristic curve of 0.888. The corresponding values for the Modified CT Severity Index were lower. At matched specificity, the AI model also showed higher sensitivity for severe and mild disease, although not every sensitivity comparison reached statistical significance. Performance did not show statistically significant differences across patient groups stratified by sex or age.
External validation showed similar performance. The AI model reached an area under the curve of 0.887 for severe acute pancreatitis and 0.858 for mild disease. It outperformed the Modified CT Severity Index and the Bedside Index of Severity in Acute Pancreatitis in this external cohort. Hybrid approaches also showed value. Combining AI with established scores produced the highest external performance in some comparisons, particularly when CT findings and clinical variables were both available. For severe disease, the combined AI, Modified CT Severity Index and Bedside Index approach produced the strongest external discrimination. For mild disease, the hybrid models also performed slightly above standalone AI. The results indicate that imaging-based AI can capture prognostic signals that may complement conventional scoring tools, rather than simply replace them.
Triage Potential and Remaining Cautions
A retrospective triage simulation classified patients as high risk, intermediate risk or low risk. In the internal test set, the AI model identified half of the patients who developed severe acute pancreatitis as high risk and identified most mild cases as low risk. Neither AI nor the Modified CT Severity Index placed any severe internal cases in the low-risk category. In the external set, the AI model identified nearly three quarters of severe cases as high risk, outperforming the Modified CT Severity Index. For low-risk triage, both methods misclassified one severe external case as low risk.
Heatmap assessment showed that the model mainly focused on features linked with severe disease, including pancreatic necrosis, peripancreatic fluid or necrotic collections, fat stranding and ascites. Misclassified cases did not show clear radiological features explaining their later clinical course. These results support caution in clinical use. High-risk outputs may indicate a need for closer vigilance rather than automatic invasive intervention. Low-risk outputs should not prevent escalation of care when clinical deterioration occurs. The model is not intended to encourage indiscriminate imaging. Its proposed role applies when early CT has already been acquired as part of patient assessment.
AI-assisted analysis of admission CT scans may support early risk stratification in acute pancreatitis when CT imaging is already performed. The model performed comparably to or better than established prognostic tools and retained performance in an external multicentre cohort. Hybrid models suggest added value when imaging and clinical scores are available together. Important limitations remain, including the retrospective design, lack of cost-effectiveness or outcome analysis, incomplete aetiology information and absence of clinical variables in the AI model. Prospective evaluation is still needed to establish clinical utility and assess real-world outcomes.
Source: Radiology Advances
Image Credit: iStock
References:
Xu Y, Teutsch B, Zeng W et al. (2026) Deep Learning-based Prediction of Acute Pancreatitis Severity From Abdominal CT With Multicenter External Validation. Radiology Advances: umag020.
