Non-Invasive Diagnosis Model for Pancreatic Cystic Tumors Based on Machine Learning-Radiomics Using Contrast-Enhanced CT

Abstract

Background: To establish and validate a radiomics diagnosis model for classification of three subtypes of the pancreatic lesion, including intraductal papillary mucinous neoplasm (IPMN), solid-pseudopapillary neoplasm (SPN) and pancreatic cystadenoma (PCN). Methods: The contrast-enhanced (CECT) images and clinical characteristics of 134 pathological proven pancreatic cystic tumor patients were retrospectively collected in this study, including IPMN (n=40), SPN (n=47) and PCN (n=47). All patients were randomly split into the training cohort (n=90) and independent validation cohort (n=44). A total of 468 radiomics features were extracted from each tumor region. The highly inter-correlated radiomics features were excluded, and Boruta algorithm was used for further feature selection. Supporting vector machine (SVM) and random forest (RF) methods were used to construct the diagnosis model based on selected features. The diagnosis overall accuracy, precision, recall, F1 score were calculated to evaluate the performance of both diagnosis models. Results: Eighty-two radiomics features with an inter-correlation coefficient of less than 0.75 were selected. Sixteen radiomics features, and eight clinical parameters, including age, gender, tumor location, tumor number, alanine aminotransferase, aspartate aminotransferase, carcinoembryonic antigen and fasting blood glucose, showed a significant difference between IPMN, SPN and PCN. Utilizing the RFE algorithm, six radiomics features and six clinical parameters were demonstrated essential for model construction. An SVM model with a linear kernel was constructed showing overall diagnosis accuracy of 82.2% in the training dataset and 81.8% in the validation dataset. The built RF model showed diagnosis overall accuracy of 100% in the training dataset and 81.8% in the validation dataset. Both SVM and RF model can identify IPMN, SPN and PCN with high diagnostic accuracy. Conclusion: Our study suggests CECT-based radiomics model can serve as a diagnostic tool for preoperative diagnosis for IPMN, SPN and PCN patients in a non-invasive, convenient and accurate way, thus facilitating evidence-based medical decision making. Funding Statement: This work was supported by the National Science and Technology Major Project of China [Grant number: 2017ZX10203205]; National Science and Technology Project of China [Grant number: 2017YFC0108704]; Natural Science Foundation of China [Grant No. 81871351]; Zhejiang Provincial Natural Science Foundation of China [LR16F010001, Y16H180003]; Zhejiang University Education Foundation ZJU-Stanford Collaboration Fund; Opening Fund of Engineering Research Center of Cognitive Healthcare of Zhejiang Province Declaration of Interests: [The authors] declare no personal conflicts of interest. Ethics Approval Statement: This retrospective study was approved by the Institutional Review Board of the First Affiliated Hospital, Zhejiang University School of Medicine. The patient consent was waived by the patients for the use of the pre-existing medical images.