CT And Guidelines Based Human-machine Fusion Model For Predicting Resectability Of The Pancreatic Cancer

Objective:To develop a Human-machine fusion learning model for predicting the resectability of Pancreatic Cancer based on CECT(contrast enhanced computed tomography)and Guidelines,which will open up a new way for the accurate treatment of the Pancreatic Cancer patients.Method:We retrospectively studied 621 pancreatic cancer patients from 4 major hospitals in recent 10 years,after removal of the multiple organ metastasized casesand nonsuitable cases,we have left 349 cases of pancreatic cancer patients.171 cases from Nanfang Hospital and 92 cases from Sun yat-sen memorial Hospital were used as the Primary Training Cohort,66 cases from Zhujiang Hospital and 20 cases from Fudan University Shanghai Cancer Center were used as the independent test dataset,The 5-fold cross validation technique was employed to obtain the better model for the prediction.In this study,Classifier fusion strategy based on evidential reasoning(CFS-ER)were used.Firstly,a multi-classifier fusion machine learning model based on CECT is constructed,and then fused with expert’s diagnosis results based on Guidelines.Among them,The multi-classifier integrated machine learning model based on CECT was composed of Conventional Radiomics(cRad)model which based on artificial features,Kernelled support tensor machine(KSTM)which has non-deep learning features,and ResNet model which has deep learning feature are fused together,so as to better mine the feature expression of the CT images.Firstly,the semi-automatic module of ITK-SNAP software was used to assist CT image segmentation to obtain 2D layer by layer tumor Region of Interest(ROI),which was then stacked into 3D ROI.Then,788 3D artificial features were extracted using Pyradiomics,and independent sample T-test was used.Then through the Least Absolute Shrinkage and Selection Operator(LASSO),and finally through the Sequential forward selection method(SFS)further filtered the features,then the features were connected to the support vector machine(SVM)to construct the classifier,so as to obtain the cRad model.Then,3D ROI was used to unify the resolution by 3D spline interpolation method,and the 3D tumor image tensor was constructed(interpolated 3D ROI was placed in the center,and zero was filled around).Using 3D tumor image tensor as input,KSTM and ResNet models were constructed respectively.Finally,based on CFS-ER,cRad,KSTM and ResNet are fused into a multi-classifier fusion machine learning model which we called CECTsignature.2 experts with more than 10 years of clinical experience were invited to re-evaluate each patient based on their CECT following the NCCN Guidelines to obtain resectable,unresectable,and critically resectable diagnoses.The three results were converted into probability values of 0.25,0.75 and 0.50 according to the traditional empirical method.Then it is used as an independent classifier and integrated with multi-classifier machine learning model to obtain a human-machine fusion learning model.Results:The multi-classifier fusion machine learning model has a sensitivity of 78.95%,specificity of 80.60%,accuracy of 80.23%,and AUC of 0.8610,which is better than cRad,KSTM,ResNet based single classifier models or their two classifier fusion models.This means that three different models,cRad,KSTM and ResNet,have mined complementary CECT feature expression from different perspectives,and can be integrated through CFS-ER,so that the fusion model has better performance.The human-machine hybrid learning model further improved the prediction performance,achieving a sensitivity of 84.21%,a specificity of 82.09%,an accuracy of 82.56%,and an AUC of 0.8845,which was better than the single NCCN-based diagnostic and multi-classifier fusion machine learning model.This means that machine learning models might learning extra information from enhanced CT that experts cannot distinguish,thus complementing expert experience and improving the performance of hybrid learning models.Conclusion:Based on CFS-ER,a multi-classifier fusion machine learning model based on CECT is firstly constructed.Then,the expert diagnosis based on NCCN Guidelines is taken as an independent classifier,which is further integrated with the multi-classifier fusion machine learning model to obtain the Human-machine fusion Machine Learning model.Independent test performance of the Human-machine fusion Machine Learning model is great.The model proposed by us is expected to be used noninvasively predict the resectability and unresectability of the pancreatic cancer before surgery in the future,so as to eliminate the border line resectable gray area of judgment,help surgeons make more accurate choices.It will help us reduce the postoperative complication rates,avoid unnecessary surgical explorations,improve individual and overall survival rates,improve patients quality of life,and lay a more solid foundation for the systematic treatment of the pancreatic cancer.