CT-based Radiomics In The Prediction Of High-and Low-risk Histopathological Classification Of Gastric Cancer

Purpose:Extract and screen the radiomic features of gastric cancer CT images,and combined with the relevant preoperative clinical features of patients,to explore the value of CT-based radiomic in predicting high and low risk histological classification of gastric cancer before operation.Method:A total of 570 patients with gastric cancer confirmed by postoperative pathology in the people's Hospital affiliated to Jiangsu University from January 2012 to December 2017 were analyzed retrospectively.The pathological results were papillary adenocarcinoma?n=7?,well/moderately differentiated tubular adenocarcinoma?n=285?,poorly differentiated adenocarcinoma?n = 247?,mucinous adenocarcinoma?n = 21?and signet ring cell carcinoma?n=10?.According to the conclusion of the literature,They were divided into two groups: high-risk histological classification group and low-risk histological classification group.All the samples were randomly divided into a training group and a validation group according to the proportion of 2:1 for model construction and verification.MSCT were performed with plain and triphasic dynamic contrast enhancement before surgery,and CT-TNM staging diagnosis was performed.The axial reconstruction thickness of plain scan,arterial phase and venous phase was 5 mm,the coronal reconstruction thickness of venous phase was 3 mm,matrix:512 × 512.The reconstructed images were uploaded to PACS with uncompressed DICOM data.The attending physician in the abdominal group of the radiology department used ITK-SNAP to segment the region of interest of gastric cancer from the CT images of the patients,and selected the maximum level of the tumor on the axial image of the CT portal vein to segment the region of interest,and then the image was corrected by the deputy chief physician and the chief physician.Multivariate analysis was used to screen out the clinical features with the greatest correlation with high and low risk histological classification of gastric cancer,and to construct a clinical characteristics model.Python-based Pyradiomics package is used to extract radiomic features from all segmented images,and the maximum correlation minimum redundancy?m RMR?,minimum absolute shrinkage and selection algorithm?LASSO?and stepwise regression were used to select the most effective features to establish a radiomic signature.Multivariate Logistic regression analysis was used to construct a nomogram that fused radiomic signature and clinical characteristics.All of the models were built on the training group and tested on the validation group.The area?AUC?under the receiver operating characteristic?ROC?curve was used to evaluate the performance of the model,and its specificity,sensitivity and accuracy were calculated.In addition,the correction curve and Hosmer-Lemeshow test were used to evaluate the robustness of radiomics models.Results:There were 278 males and 103 females in the training group and 133 males and 56 females in the verification group.The results of univariate analysis showed that there was no significant difference in the characteristic distribution between the two groups.The results of multivariate logistic regression analysis showed that age,sex and CT-M stage were significantly correlated with different risk histological classification of gastric cancer,so these characteristics were included in the clinical characteristics model.985 radiomic features were extracted from each regions of interest?ROI?,including 18 intensity features,12 shape features,75 texture features and 880 wavelet features.After the screening of redundant features,five features which were most related to high and low risk histological classification of gastric cancer were obtained:“Wavelet.HL_glszm_Zone Percentage”,“Wavelet.LH_glszm_Gray Level Non Uniformity”,“Wavelet.LH_glszm_Zone Variance”,“Wavelet.HH_glrlm_Low Gray Level Run Emphasis”,“Wavelet.LL_Firstorder₉0Percentile”.The AUC of clinical characteristic model in training group and verification group was60.32%?95%CI:0.546-0.660?and 61.26%?95%CI:0.530-0.694?respectively,and its specificity,sensitivity and accuracy in training group and verification group were 0.6613,0.5282,0.5932 and 0.7264,0.4337,0.5979,respectively.The AUC of radiomic signature in the training group and verification group were 67.60%?95%CI:0.622-0.729?and 65.20%?95%CI:0.572-0.731?,respectively.The specificity,sensitivity and accuracy in the two groups were 0.8710,0.4205,0.6504 and 0.8113,0.4458,0.6508,respectively.The AUC of CT-based radiomics model combining preoperative clinical and radiomic features was 70.69%?95%CI:0.655-0.758?in the training group and 69.49%?95%CI:0.618-0.771?in the verification group.The specificity,sensitivity and accuracy of the two groups were 0.7634,0.5487,0.6535 and 0.7736,0.5181,0.6614,respectively.The performance of radiomics model in predicting the high and low risk histological classification of gastric cancer is the best among the three models,followed by radiomic signature,and the worst is the clinical characteristic model.The calibration curve and Hosmer-Lemesshow test results?paired 0.5627?showed that CT-based nomogram fit well between prediction and the clinical observation results.Conclusions:1.Radiomic signature is more effective in predicting the high and low risk histological classification of gastric cancer than the preoperative clinical characteristic model.2.The diagnostic efficiency of the joint predictive model combining clinical and radiomic features is higher than that of a single predictive model.3.CT-based radiomic model has a certain reference value in non-invasive prediction of high and low risk histological classification of gastric cancer before operation.