Texture Analysis As A Radiomic Marker For Differentiating Pancreatic Ductal Adenocarcinoma,Solid Pseudopapillary Tumor Of Pancreas And Pancreatic Neuroendocrine Tumor

Objective: To differentiate diagnostic pancreatic ductal adenocarcinoma,pancreatic neuroendocrine tumors and solid pseudopapillary tumors of the pancreas and to retrospective evaluate the diagnostic performance based on contrast CT texture analysis and machine learning algorithms.Materials and Methods: This retrospective study included a total of 199 cases pathologically confirmed pancreatic tumors,including 98 cases of pathologically confirmed pancreatic ductal adenocarcinoma,62 cases of solid pancreatic pseudopapillary tumors and 39 cases of pancreatic neuroendocrine tumors,all patients underwent pancreatic contrast CT scan.Tumor boundaries were manually sketched in the images of the pancreas AP and 46 texture features were extracted,including 13 histogram features,14 gray level co-occurrence matrix features,8 gray-scale run length matrix features and 11 gray-level area size matrix features.In univariate analysis,texture features of different pancreas tumors were compared separately,and the area under the curve(AUC)was calculated.In multivariate analysis,the random forest method was used to select texture features.Pancreas tumors subtype was classified by six different machine learning methods(Linear Discriminant Analysis,K Nearest Neighbor,Random Forest,Adaboost,Naive Bayes,Neural Network).The selected features were classified the tumors type by those machine learning models,and calculate the area under the curve(AUC)based on 10-fold cross-validation.AUCs were classified as follows: 0.50–0.59,poor;0.60–0.69,fair;0.70–0.79,good;0.80–0.89,very good;0.90–1.0,excellent.Results: In univariate analysis,Low Intensity Emphasis,Low Intensity Small Area Emphasis and Low Intensity Large Area Emphasis showed good discrimination ability for discriminating pancreatic ductal adenocarcinoma,and the best feature was Low Intensity Small Area Emphasis with AUC of 0.7345(P <0.0001).Short Run Emphasis,Grey Level Nonuniformity and Run Length Nonuniformity showed good discriminating ability for identifying SPT,and the best feature was Grey Level Nonuniformity with AUC of 0.7900(P <0.0001).The sum average shows a very good discriminating ability for the identification of PNETs with AUC of 0.8930(P <0.0001).In multivariate analysis,features were selected based on random forest algorithm,the maximum AUC of six machine learning algorithms for discriminating PDAC,SPT and PNETs were 0.8819(random forest)and 0.8583(random forest)and 0.9390(Adaboost),respectively and discrimination ability has reached very good or excellent degree.The six machine learning algorithms classify tumors directly,the classification accuracy rate reached 80%(random forest algorithm).Conclusion: Non-invasive identification of PDAC,SPT and PNETs based on texture analysis of contrast CT is reliable,and using machine learning algorithm can further improve the performance of differential diagnosis.