CT-Based Radiomics And Deep Learning For Pretreatment Prediction Of Prognosis In Advanced Gastric Cancer

Purpose: Gastric cancer is one of the most common malignant tumors of the digestive tract.Predicting the probability of survival and early recurrence accurately before surgery,stratifying the risk,are the key steps to individualized treatment of gastric cancer.This study aimed at the clinical needs of preoperative prediction of OS and early recurrence of advanced gastric cancer,and constructed an radiomics and deep learning prediction model based on enhanced CT.Materials and methods: 1.We retrospectively collected 337 patients from Lanzhou University Second Hospital,which were divided into a training cohort(n=237)and an external validation cohort(n=100).We developed an end-to-end DL model based on the architecture of residual convolutional neural network,and augmented the size of training dataset by image transformations to avoid overfitting.Cox regression for univariable and multivariable analysis were used to develop a clinical model.Meanwile,a comprehensive prediction model was constructed by combining the deep learning model and the clinical model,and then,nomogram was constructed.We calculated the Harrell's concordance index(C-index)and hazard ratio(HR)to evaluate the performance of the three models.Calibration curves and decision curve analysis(DCA)was also applied to verify the prognostic value of these prediction models 2.We enrolled 669 consecutive patients(302 in the training set,219 in the internal test set and 148 in the external test set)with clinicopathologically confirmed AGC from two centers.Radiomic features were extracted from preoperative diagnostic CT images.Additionally,a deep convolutional neural networks(DCNN)was constructed and trained to extract deep learning features.Machine learning methods were applied to shrink feature size and build a predictive radiomic signature.We incorporated the radiomic signature and clinical risk factors into a nomogram using multivariable logistic regression analysis.The area under the curve(AUC)of operating characteristics(ROC),accuracy,stratified analysis and calibration curves were assessed to evaluate the nomogram's performance in discriminating early recurrence.Results: 1.Totally 5688 CT images were prepared by data augmentation and fed into DL model.The trained DL model significantly classified patients into high-risk and low-risk groups in training cohort(concordance index(C-index): 0.70,hazard ratio(HR): 2.88,P-value<0.002)and external validation cohort(C-index:0.64,HR: 4.32,Pvalue<0.002,).The clinical model was developed with three significant clinical variables(P-value<0.05).The comparison illustrated comprehensive prediction model(DL+ Clinical)had the best performance for risk prediction of OS according to the Cindex(training: comprehensive prediction model vs DL vs Clinical =0.74 vs 0.72 vs 0.70;external validation: 0.67 vs 066 vs 0.64).The DL model had the highest HR both in training cohort:(DL vs Clinical vs Comprehensive prediction model =2.88 vs 2.72 vs 2.72;external validation cohort: 4.32 vs 2.11 vs 1.89),which indicated the high-risk groups predicted by DL model had higher hazard of death than the high-risk groups predicted by other models.The nomogram,calibrations and DCA curves showed that compared with DL and clinical model,the comprehensive prediction model has the best performance for risk prediction.The calibration curve analysis shows that the predicted 1,2 and 3 year survival probability based on the integrated prediction model is in good agreement with the actual OS.The DCA also proved that the comprehensive prediction model had the best net benefit than other models for the patients.2.A radiomic signature,including three hand crafted features and six deep learning features,was significantly associated with early recurrence(p-value <0.0001 for all sets).In addition,clinical N stage,carbohydrate antigen 199 levels,carcinoembryonic antigen levels,and Borrmann type were considered useful predictors for early recurrence.The nomogram,combining all these predictors,showed powerful prognostic ability in the training set and two test sets with AUCs of 0.831(95% CI,0.786–0.876),0.826(0.772–0.880)and 0.806(0.732–0.881),respectively.The predicted risk yielded good agreement with the observed recurrence probability.The stratified analysis indicated that the performance of radiomic nomogram was not affected by different subgroups(DeLong test p-value > 0.05),which implied its great generalization ability.The calibration curve of the radiomic nomogram showed good agreement between the predictive risk and the observed recurrence probability in the three sets.The Hosmer-Lemeshow test was not significant(p= 0.4667,0.1372 and 0.0937),suggesting there is no significant departure.Conclusion: 1.The end-to-end DL model constructed based on the residual convolutional neural network in this study is a good survival risk assessment model and has good application value for the early prediction of OS in AGC patients.2.The combination of the advantages of radiomics and DL,and the fusion of clinical independent risk factors to construct radiomics nomogram can predict the early recurrence probability of AGC patients before surgery.This may become a potential tool to guide the personalized treatment of AGC.The new AI technology represented by radiomics and DL is expected to effectively supplement the existing prognostic evaluation system.