Machine learning based texture analysis predicts postoperative pancreatic fistula in preoperative non-contrast enhanced computed tomography

Background: Postoperative pancreatic fistula (POPF) remains an unsolved challenge after pancreaticoduodenectomy (PD) and soft pancreatic texture represents a major risk factor for POPF. Technical advances in machine learning and artificial intelligence extend the possibilities of texture analysis (TA). This study investigates the potential of machine learning based approaches to describe pancreatic texture and to predict POPF based on non-contrast enhanced Computed Tomography (CT). Material & Methods: A prospectively assessed database including all patients undergoing PD at a tertiary center, was screened for 110 patients based on the occurrence of POPF. Group A consisted of 55 patients without POPF and group B of 55 patients who suffered from POPF. Based on recognized risk factors of POPF, the validated original fistula risk score (FRS) and the alternative FRS (www.pancreasclub.com) were utilized. For machine learning based TA preoperative non-contrast-enhanced low radiation dose CT image data and WEKA© ML software were used. Regions-of-interest (ROI) for TA encompassing the pancreatic stump were drawn free hand.The potential of ML derived texture features to predict intraoperative histological characteristic, hardness of pancreas and POPF was analyzed. ML results were tested using ten- fold cross validation. Results: Both the original FRS and the alternative FRS showed good discrimination between patients without (Group A) and with POPF (Group B) with an area under the curve (AUC) of 0.76 and 0.72 respectively. ML analysis revealed a potential to predict histological fibrosis (AUC 0.84, sensitivity 75%; specificity 92%), lipomatosis (AUC 0.82, sensitivity 78%; specificity 89%) and intraoperative pancreatic hardness (AUC 0.70, sensitivity 78%; specificity 74%). Additionally, ML texture features were most accurate in predicting the occurrence of POPF (AUC 0.95, sensitivity of 96%; specificity 98%). Conclusion: This study indicates the ability of ML algorithms in recognizing novel pancreatic texture features and predicting POPF based on easily accessible preoperative imaging.