Study Of Prediction Algorithm Of Genetic Mutation Of GISTs Based On Radiomics And CNN

Gastrointestinal stromal tumors(GISTs)are the most common stromal tumors in the gastrointestinal tract.They are mainly found in elderly patients,with an incidence ranging from 11 to 19.6 per million people worldwide.The study found that most gastrointestinal stromal tumors are driven by oncogenic mutations of the proto-oncogene KIT or PDGFRA,while imatinib mesylate can inhibits proto-oncogenes as the targeted therapy.In order to determine the genotype of a patient,invasive surgery is required to extract and analyze the tumor,which is not only difficult but also has risks.In order to non-invasively predict the genetic mutation of patients with GISTs,this paper proposes a prediction algorithm for gastrointestinal stromal tumor gene mutation based on Radiomics and convolutional neural network(CNN).The algorithm extracts the global and local features of CT images of GISTs by Radiomics and convolutional neural network,network.And then these features are concatenated as combined features.After feature selection,the logistic regression model was used to classify the variation of the PDGFRA_EXON18 gene with combined features.This paper mainly completed the following work:1)Extracting 1,803 Radiomics features from CT images using PyRadiomics in ten image types and six feature systems.2)Extracting 1083 CNN features using the VGG-16 pre-trained with ImageNet.3)Using Mann-Whitney U test to remove features that are not significantly different from the positive and negative conditions of the classification and using LASSO regression to select features.4)Construct logistic regression models to predict the genetic mutation of PDGFRA EXON 18 based on Radiomics features,CNN features and combines features.The experimental results show that among the three characteristics,the classification model of combined feature training has the best effect(Accuracy,0.797[95%CI,0.738 to 0.855],AUC,0.888[95%CI,0.854 to 0.922],Sensitivity,0.730[95%Cl,0.636 to 0.824],Specificity,0.817[95%CI,0.731 to 0.902]).The algorithm proposed in this paper is not effective in sensitivity,but the comprehensive index AUC has reached a high level.It means that the algorithm has potential application value in predicting genetic mutation of PDGFRA_EXON18 of patients with GISTs.In addition,based on the algorithm proposed in this paper,non-invasive detection of gene mutations can be achieved,and postoperative infections can be avoided to assist doctors in diagnosis and treatment.