Assessment of machine learning in radiomics for predicting the risk of clinically significant pancreatic fistulas after pancreatoduodenal resections using CT imaging
Open Access
- 28 November 2025
- journal article
- Published by Vidar, Ltd. in Medical Visualization
- Vol. 29 (4) , 81-93
Abstract
Despite the reduction in mortality after pancreatoduodenal resections, the incidence of postoperative complications remains high (20–60%). One of the most severe complications is postoperative pancreatic fistula. Predicting the risks of a complicated postoperative period remains an urgent problem. One of the key risk factors is the pancreatic structure, and modern approaches to predicting clinically relevant pancreatic fistula integrate not only morphological but also radiomic parameters of CT images using artificial intelligence methods. Aim. To evaluate the potential of machine learning in radiomics for predicting clinically relevant pancreatic fistulas after pancreatoduodenal resection and to develop a clinical decision support system based on the “Virtual Biopsy” platform. Methods. Retrospective analysis of data from 117 patients who underwent pancreatoduodenal resection (2016–2019) at the A.V. Vishnevsky National Medical Research Center of Surgery. Machine learning methods were applied to assess textural features of preoperative CT scans. Results. Сlinically relevant pancreatic fistulas were recorded in 31 patients (26.5%). Clinically significant fistula alone was diagnosed in 11 patients (9.4%), while its combination with arrosive bleeding was observed in 20 patients (17.1%). The peak incidence of fistulas occurred on days 4–6, and of bleeding on days 8–14. The radiomic AdaBoost model demonstrated the highest efficacy (ROC AUC = 0.815), outperforming alternative approaches: Gradient Boosting (0.631), XGBoost (0.677), LightGBM (0.631), and Stacking (0.662). Integration of morphological features did not improve predictive capability, likely due to data noise. Models based on semantic parameters (max. ROC AUC = 0.653) confirmed limited clinical applicability. Conclusion. Machine learning methods are effective in predicting clinically relevant pancreatic fistulas after pancreatoduodenal resections. Radiomic analysis extends the diagnostic potential of CT, demonstrating superior model accuracy metrics compared to classical semantic features alone.Keywords
This publication has 32 references indexed in Scilit:
- Preoperative prediction of clinically relevant postoperative pancreatic fistula after pancreaticoduodenectomyEuropean Journal of Radiology, 2021
- Computed tomography-adjusted fistula risk score for predicting clinically relevant postoperative pancreatic fistula after pancreatoduodenectomy: Training and external validation of model upgradeEBioMedicine, 2020
- Elucidating early CT after pancreatico-duodenectomy: a primer for radiologistsInsights into Imaging, 2018
- Major pancreatic resections: normal postoperative findings and complicationsInsights into Imaging, 2018
- Association Between the Incidence of Pancreatic Fistula After Pancreaticoduodenectomy and the Degree of Pancreatic FibrosisJournal of Gastrointestinal Surgery, 2018
- Are there preoperative factors related to a “soft pancreas” and are they predictive of pancreatic fistulas after pancreatic resection?Surgery Today, 2014
- Pre-operative prediction of pancreatic fistula: Is it possible?Pancreatology, 2013
- 3D Slicer as an image computing platform for the Quantitative Imaging NetworkMagnetic Resonance Imaging, 2012
- Dual-Phase Computed Tomography for Assessment of Pancreatic Fibrosis and Anastomotic Failure Risk Following PancreatoduodenectomyJournal of Gastrointestinal Surgery, 2011
- Postoperative acute pancreatitis as a major determinant of postoperative delayed gastric emptying after pancreaticoduodenectomyJournal of Gastrointestinal Surgery, 2006