Objective: Gastrointestinal stromal tumors(Gastrointestinal stromal tumors,GISTs)with certain malignant potential,are the most common gastric submucosal tumors.it is significant to predict the malignant potential of gastric GISTs accurately before operation.This study aimed to construct a deep learning model to predict the risk stratification of gastric GISTs by identifying endoscopic ultrasound(EUS)images;construct a traditional risk prediction model by using clinical data and EUS features,then combine the two prediction models to form a third joint prediction model,to compare the predictive efficacy of the above three models.Methods: A total of 923 images and clinical data of 310 gastric GISTs confirmed by pathological examination were included.Firstly,the EUS images of different risk stratification were partitioned into 3 groups randomly with 70% for training the deep learning model,20% for testing and adjusting the trained model,and10% inputting model to obtain the risk stratification result of the model prediction.We also included a subgroup to explore the efficacy of the deep learning model in predicting the ki-67 index.Meanwhile,we performed univariate analysis of patients’ clinical data and EUS features,and factors with statistically significant differences(P < 0.05)were included in the multifactorial analysis to establish a traditional risk prediction model.Then,we combined the above two models to construct a joint prediction model.We compared the predictive efficacy of the three models by identifying 96 EUS images of 31 patients which were selected from the test set of the deep learning model randomly,plotted their ROC,and compared their area under the curve(AUC).Results : The average sensitivity,specificity,positive predictive value,negative predictive value and accuracy of the deep learning model in the test set for predicting the risk stratification were 82.45%、94.57%、84.78%、94.98% and 85.29%,respectively.The average sensitivity,specificity,positive predictive value,negative predictive value,and accuracy of the deep learning model for predicting the range of Ki-67 index by recognizing the ultrasound endoscopic images were 90.06%,94.52%,89.37%,94.11%,and 89.86%,respectively.During the construction of the traditional model,there were statistically significant differences in the location,growth pattern,border,internal echo,and rupture between the different risk gastric GISTs in EUS features(p <0.05)by univariate analysis.Multivariate logistic regression analysis showed that the maximum size and internal echo were independent factors for the identification of gastric GISTs with different risk stratification(OR were 3.03,0.522 and P were P <0.001 and 0.026)respectively.The average sensitivity,specificity,positive predictive value,negative predictive value,and accuracy of the traditional risk prediction model built by Logistic regression were 67.05%,94.29%,73.11%,90.01%,and 78.71%,respectively.In comparing the predictive efficacy of the three models for malignant potential of gastric GISTs,the combined model had the highest accuracy of 90.32%,followed by the deep learning model with 87.88% accuracy,while the traditional model had 71.10% accuracy.The AUC of the combined model was 0.932,which was better than the traditional prediction model at 0.909,and the AUC of the deep learning model was 0.843,but the difference was not statistically significant when the AUCs of the three models were compared between the three groups.Conclusion : 1.The deep learning model had great potential for predicting the risk stratification and the ki-67 index of gastric GISTs;2.In this study,the maximum diameter of lesion,growth mode,site,boundary,internal echo and rupture were all influencing factors of gastric GISTs risk stratification,among which the maximum diameter of lesion and internal echo were independent influencing factors of gastric GISTs risk stratification;3.The accuracy and prediction performance of the combined model for gastric GISTs risk prediction were both higher than those of the deep learning model or the traditional risk prediction model alone,but the differences were not statistically significant. |