Preliminary Investigation Of Gastric Cancer Radiomics Enhanced CT For Predicting Synchronous Liver Metastasis

Purpose:The purpose of this study is to predict synchronous liver metastases using radiomics analysis on preoperative contrast-enhanced CT images of gastric cancer.Materials and methods:This study is a retrospective,single-center investigation that was conducted from June 2016 to June 2022.The study included patients with both synchronous and nonsynchronous liver metastases who underwent preoperative contrast-enhanced CT scans for gastric cancer.The collected data consisted of the arterial phase,venous phase,and delayed phase enhanced CT images.Age,gender,pathological T,pathological N,M,and liver metastasis were within the information collected.The training set(7:3 ratio)and test set were created from the collected case data.3D Slicer was used to manually contour the arterial phase,venous phase,and delayed phase images,with the largest axial slice of the gastric cancer as the region of interest(ROI).The ROI was tracked by one radiologist and reviewed by another,with both of them blinded to the patient’s clinical information.Radiomics features were extracted and normalized using Python.The Mann-Whitney U test was used for feature selection and only features with pvalues less than 0.05 were retained.One feature with high correlation was retained using the Spearman test.The key features for prediction were obtained using the least absolute shrinkage and selection operator(LASSO)and cross-validation.Multiple machine learning methods are used to build prediction models.Multiple statistical methods were utilized to evaluate the efficacy of the predictive model.Results:The retrospective study included 76 patients,33 in the synchronous liver metastasis group and 43 in the non-synchronous liver metastasis group.The outcomes showed significant differences in pathological N stage(p=0.006),M stage(p<0.001),gender(p=0.031)and age(p=0.02).The pathological T stage,however,showed no meaningful difference(p>0.05).The distribution of these features in the training and test sets was found to be statistically similar(P>0.05).Except for the M stage,the distribution of these features before surgery,during surgery,or within 6 months postsurgery in the synchronous liver metastasis group showed no statistically significant difference(P>0.05).A total of 851 radiomics features were extracted for each phase(arterial,venous,and delayed)for each case.A total of 1702 radiomics features were obtained by merging the data from the arterial and venous phases.The radiomics features of the arterial phase,venous phase,delayed phase,and the combination of arterial and venous phases were selected to 3,9,3,and 9 features,respectively.Based on each of these feature classes,prediction models were built using six machine learning methods.The logistic regression(LR)model in the venous phase and the LR and support vector machine(SVM)models in the combination features from both arterial and venous phases performed well.The LR model in the venous phase had an AUC of 0.90(95% CI 0.78-1.00),accuracy of 0.87,sensitivity of 0.78,specificity of0.93,positive predictive value of 0.87,negative predictive value of 0.87,precision of0.87,recall of 0.78,F1 score of 0.82,area under precision recall curve(AUPRC)of0.85,and Brier score of 0.14.The LR model in the combination features from both arterial and venous phases had an AUC of 0.97(95% CI 0.90-1.00),accuracy of 0.91,sensitivity of 0.89,specificity of 0.93,positive predictive value of 0.89,negative predictive value of 0.93,precision of 0.89,recall of 0.89,F1 score of 0.89,AUPRC of0.96,and Brier score of 0.14.In the combination features from both arterial and venous phases,the training set SVM model had a higher accuracy of 0.81 compared to the LR model with an accuracy of 0.70.The SVM model in the test set had a lower performance compared to the LR model but still performed well with an AUC of 0.89(95% CI 0.73-1.00),accuracy of 0.83,sensitivity of 0.67,specificity of 0.93,positive predictive value of 0.86,negative predictive value of 0.81,precision of 0.86,recall of0.67,F1 score of 0.75,AUPRC of 0.88,and Brier score of 0.20.The decision curve analysis suggests that the above-mentioned model has some clinical utility.Conclusion:Radiomics based on enhanced CT in gastric cancer can predict the risk of synchronous liver metastasis in gastric cancer.The combined arterial and venous phase LR models,as well as the venous phase LR model,and the combined arterial and venous phase SVM model,demonstrated good performance in prediction with high accuracy.The research results indicate that the established models have clinical utility.