Development Of Preoperative Prediction Model Of Lauren Classification In Gastric Cancer Based On CT Radiomics

ObjectivesBased on the study of CT image,we analyzed the difference of pathological characteristics between intestinal type and diffuse type,and develop a radiomic nomogram to preoperatively differentiate Lauren diffuse type from intestinal type in gastric cancer(GC).MethodsThis study was approved by the ethics review committee of Zhenjiang first people's hospital.A dataset of 539 GC patients were totally enrolled and randomly separated into two cohorts at a 7:3 ratio for training and validation.All patients received preoperative enhanced CT and postoperative pathological examination between December 2011 and December 2017 in our study.We used ITK-SNAP software for the largest manual segmentation of the tumor habitat in the CT images.For the tumor ROIs,two radiologists who had 10 years of experience reviewed all the CT slices of a patient and selected one with the largest tumor area to segment,they subsequently delineated the tumor ROIs for all the 539 patients,and repeated the segmentation procedure after one month on 30 randomly selected patients for feature stability.All segmentations were finally approved by a senior radiologist who had 30 years of experience.A total of 2074(1037+1037)radiomic features were initially extracted from tumor ROIs and peripheral rings on each CT image,respectively,including shape features,first-order features,and texture features.First,numerical radiomic features were first standardized by z-score method using the mean and standard deviation parameters calculated from patients,then with an ICC>0.75 as a reliability standard.For simplification and generalization,we only adopted the least absolute shrinkage and selection operator(LASSO)method with10-fold cross-validation,which is commonly used to select the most valuable feature.At the same time,the clinical factor of patients(age,sex,CT_T,CT_N)was collected.All statistical analysis was performed on R software.In univariate analysis,Mann-Whitney U test was adopted for continuous clinical factor(age),and Chi-squared test or Fisher exact test were applied for categorical variables(sex,CT_T,and CT_N).A two-tailed p value<0.05 represented a statistical significance.six models(Radiomic nomogram,Combined radiomic signature,Tumor-based model,Peripheral ring-based model,Clinical model 1(CT_T+CT_N),Clinical model 2(age+CT_T+CT_N)were constructed to preoperatively differentiate intestinal and diffuse gastric cancer.To assess the discrimination ability,receiver operating characteristic(ROC)curves as well as corresponding area under ROC curves(AUC)were given for the six models.Accuracy,specificity,and sensitivity results were also attached.We adopted Delong-test to compare the predictive performance between each two models.To verify the good fitness of model predictive outputs with actual values,calibration curves were conducted for the radiomic nomogram.To quantify the usefulness in clinical trials,decision curves were conducted in the validation cohort by calculating the net benefits at some threshold probabilities.ResultsA total of 2074(1037+1037)radiomic features were initially extracted from tumor ROIs and peripheral rings on each CT image,respectively.With an ICC>0.75as a reliability standard,232 tumor-based features and 206 peripheral ring-based features were included for the subsequent feature selection.LASSO method identified three potential features,including two first-order features from tumor ROIs and one shape feature from peripheral rings.The three potential features were chosen to build a radiomics signature.Three clinical factors identified by univariate analysis(age,CT_T,and CT_N)along with the combined radiomic signature constructed the radiomic nomogram.The combined radiomic signature achieving an area under receiver operating characteristic curve(AUC)of 0.715(95%confidence interval[CI],0.663-0.767)in the training cohort and 0.714(95%CI,0.636-0.792)in the validation cohort.The radiomic nomogram incorporating the combined radiomic signature,age,CT T stage,and CT N stage outperformed the other models with a training AUC of0.745(95%CI,0.696-0.795)and a validation AUC of 0.758(95%CI,0.685-0.831).Delong-test indicated that there were significant differences between predictive performance of radiomic nomogram and that of clinical model 1,clinical model 2,tumor-based model,and peripheral ring-based model in both cohorts.As to the association between radiomic nomogram and the combined radiomic signature,an IDI of 5.71%in the validation cohort demonstrated the improvement from the combined radiomic signature to radiomic nomogram.Calibration curves for the radiomic nomogram showed good fitness.Decision curve analysis in the validation cohort indicated that radiomic nomogram added more benefits when directing treatment strategies for GC patients with different Lauren classification types,compared with the Clinical model 1 and Clinical model 2.ConclusionsThe radiomic nomogram involving the combined radiomic signature and clinical characteristics held potential in differentiating Lauren diffuse type from intestinal type for reasonable clinical treatment strategy.