The potential of machine learning to predict postoperative pancreatic fistula based on preoperative, non-contrast-enhanced CT: A proof-of-principle study

• Published in: Surgery 2020-02, Vol.167 (2), p.448-454
• Main Authors: Kambakamba, Patryk, Mannil, Manoj, Herrera, Paola E., Müller, Philip C., Kuemmerli, Christoph, Linecker, Michael, von Spiczak, Jochen, Hüllner, Martin W., Raptis, Dimitri A., Petrowsky, Henrik, Clavien, Pierre-Alain, Alkadhi, Hatem
• Format: Article
• Language: English
• Subjects: Aged; Female; Humans; Machine Learning; Male; Middle Aged; Pancreas - diagnostic imaging; Pancreatic Fistula; Pancreaticoduodenectomy; Positron Emission Tomography Computed Tomography; Postoperative Complications; Proof of Concept Study; Retrospective Studies
• ISSN: 0039-6060; 1532-7361
• DOI: 10.1016/j.surg.2019.09.019

Postoperative pancreatic fistula remains an unsolved challenge after pancreatoduodenectomy. Important in this regard is the presence of a soft pancreatic texture which is a major risk factor. Advances in machine learning and texture analysis of medical images allow identification of features of parenchyma that are invisible to the human eye. The aim of this study was to investigate the potential of machine learning to predict postoperative pancreatic fistula based on preoperative, non-contrast-enhanced computed tomography. We screened a prospectively assessed database including all patients undergoing pancreatoduodenectomy at a tertiary center from 2008 until 2018 for patients based on the occurrence of postoperative pancreatic fistula. In total, 110 patients were included, consisting of 55 patients who developed a postoperative pancreatic fistula and 55 without postoperative pancreatic fistula. For machine learning-based texture analysis preoperative, non-contrast-enhanced computed tomography axial images were used. Machine learning results were tested using 10-fold cross validation. Previously validated clinical fistula risk scores (original and alternative fistula risk scores) served as reference tests. Both the original and the alternative fistula risk scores showed good discrimination between patients without and with postoperative pancreatic fistula (area under the curve 0.76 and 0.72, respectively). Machine learning-based texture analysis showed potential to detect histologic fibrosis (area under the curve 0.84, sensitivity 75%; specificity 92%), histologic lipomatosis (area under the curve 0.82, sensitivity 78%; specificity 89%), and intraoperative pancreatic hardness (area under the curve 0.70, sensitivity 78%; specificity 74%). The features of the machine learning-based texture analysis were most accurate in predicting the occurrence of postoperative pancreatic fistula (area under the curve 0.95, sensitivity of 96%; specificity 98%) after pancreatoduodenectomy. This proof-of-principle study suggests the ability of machine learning in recognizing important features of pancreatic texture associated with an increased risk of postoperative pancreatic fistula based on preoperative computed tomography.