Study On CT Radiomics To Predict Lymphovascular Invasion Status Of Esophageal Carcinoma

Part One A preliminary study of radiomics based on contrast-enhancedCT to predict the lymphovascular invasion status in esophag-eal squamous cell carcinomaObjective:This study attempted to use the radiomics approach based on the thin-section contrast-enhanced CT（CECT）images of esophageal squamous cell carcinoma（ESCC）patients to develop a predictive model to predict the lymphovascular invasion（LVI）status of ESCC patients preoperatively and provide decision support for clinicians to develop individualized treatment plans.Methods:This study retrospectively included 334 patients with ESCC who underwent radical resection and were confirmed by postoperative pathology at our institution between August 2016 and October 2019,including96 LVI-positive patients and 238 LVI-negative patients.All enrolled patients were randomly divided into a training set and a testing set at the ratio of 7:3.The training set contained 234 ESCC patients,including 68 LVI-positive patients and 166 LVI-negative patients;the testing set contained 100 ESCC patients,including 28 LVI-positive patients and 72 LVI-negative patients.All patients had a chest or chest plus abdomen contrast-enhanced CT scan within2 weeks before surgery.The 3D Slicer software was used to outline the tumor3D region of interest layer by layer,and the Py Radiomics software package was used to extract the radiomics features of tumor tissues,and the Least absolute shrinkage and selection operator（LASSO）method was applied to filter the radiomics features.Three classification machine learning algorithms,logistic regression（LR）,support vector machine（SVM）and decision tree（DT）,were used to build the imaging histology model to predict the LVI status of ESCC patients.The performance of the radiomics models was compared,and the best radiomics model was selected to calculate Radscore score.The maximum tumor thickness(CT_Thick)was measured on thin-section enhanced CT images,and the patients were evaluated for clinical T stage（c T stage）,N stage（c N stage）and AJCC stage（c AJCC stage）.The CEA and SCCA results of patients within 2 weeks before surgery was reviewed and recorded using the hospital information system（HIS）,and the above clinical characteristics were included in the univariable logistic regression analysis.The independent predictors of LVI status were then obtained using multivariable logistic regression,and clinical prediction models were developed.The independent predictors from the clinical features were combined with the radiomics features,and a combined model was developed to predict LVI status in ESCC patients.A nomogram was created based on the results of the combined model in the training set to demonstrate the risk probability of developing LVI in individual patients.Receiver operating characteristic（ROC）and decision curves analysis（DCA）were used to evaluate the performance of each model in predicting LVI status in ESCC patients.Results:Among the radiomics models,the highest prediction performance was achieved using LR and SVM,with the area under the curve（AUC）of 0.847 in the training set and 0.826 in the test set.Sphericity and gray level nonuniformity（GLNU）were the most important radiomics features for predicting LVI status in ESCC patients.In the clinical model,CT_Thickwas significantly greater in the LVI-positive group than in the LVI-negative group（P<0.001）.The c N staging was significantly higher in LVI-positive patients than in LVI-negative patients（P<0.001）.The results of ROC analysis showed that the radiomics model（AUC values of 0.847 and 0.826 in the training and test sets,respectively）and the combined model（AUC values of 0.876 and0.867）had higher AUC values than the clinical model（0.775 and 0.798,respectively）,with the combined having the highest AUC values.Conclusions:1.The radiomics model based on arterial phase enhanced CT images was effective in predicting the LVI state of ESCC patients.2.Among the clinical features,CT_Thick,c N stage and SCCA levels were associated with predicting LVI status in ESCC patients.Multivariable analysis showed that CT_Thickand c N stage were independent predictors of LVI status in ESCC patients.3.The combined model built with imaging histological features and clinical features had the highest predictive performance and was effective in predicting the LVI status of ESCC patients,and the nomogram built based on the combined model was able to assess the risk probability of LVI in independent individuals with ESCC.Part Two A preliminary study of contrast-enhanced CT derived tumorimaging features to predict lymphovascular invasion status in squamous esophageal carcinomaObjective:To initially investigate the value of contrast-enhanced CT（CECT）image-derived tumor imaging features in predicting lymphovascular infiltration（LVI）status in esophageal squamous cell carcinoma（ESCC）.Methods:This study retrospectively included 197 patients with ESCC treated with radical surgery at our institution between January 2017 and January 2019,all of whom were confirmed by postoperative pathology,including 59 LVI-positive patients and 138 LVI-negative patients.Thoracic arterial phase thin-section enhanced CT images of all patients were retrospectively analyzed independently by two imaging physicians.The tumor imaging features were divided into CT value-related features and morphology-related features.CT value-related features included CT value of tumor tissue(CT_Tumor),CT value of normal esophagus(CT_Normal),ratio between CT value of tumor tissue and normal esophagus(CT_Ratio),and difference between CT value of tumor tissue and normal esophagus(CT_Difference).In addition,thickened small vessels within the tumor and intratumor necrosis were included.Morphology-related features included CT measurement of maximum tumor thickness(CT_Thick),CT measurement of maximum tumor length(CT_Length),CT measurement of tumor volume(CT_Volume),and tumor margin morphology.The differences of CT features in the LVI-positive and LVI-negative groups were first compared,and then the features with P<0.05 were included in the one-way logistic regression analysis to screen out the imaging features associated with LVI status.The imaging features with P<0.05 in the one-way analysis were then included in the multivariable logistic regression analysis,and the independent predictors predicting LVI status were screened out.The predictive performance of each independent predictor and its combined model was analyzed using ROC curves,and a nomogram was plotted for scoring the likelihood of LVI in independent individuals.Results:Among the CT value-related features,CT_Tumor,CT_Ratioand CT_Differencewere significantly higher in the LVI-positive group than in the LVI-negative group（P<0.001,P=0.001,P<0.001）.The difference between CT_Normalin the LVI-positive and negative groups was not statistically significant（P=0.413）.In the LVI-positive group,the proportion of tumor heterogeneous enhancement（28.8%）,intratumor thickened small vessels（61.0%）,and intratumor necrosis（28.8%）was significantly higher than that in the LVI-negative group（9.42%,29.0%,12.3%;all P<0.05）.Among morphology-related features,CT_Thick,CT_Length,and CT_Volumewere significantly higher in the LVI-positive group than in the LVI-negative group（P<0.001,P<0.001,P<0.001）.In addition,the proportion of tumor margin irregularities was significantly higher in the LVI-positive group（57.6%）than in the LVI-negative group（15.2%;P<0.001）.A univariable logistic regression analysis showed that CT_Tumor,CT_Ratio,CT_Difference,tumor heterogeneous enhancement in CT value-related features and CT_Thick,CT_Length,CT_Volume,and tumor margin morphology in morphology-related features were all associated with LVI status（P<0.05）.The multivariable logistic regression analysis showed that CT_Ratio（OR,8.655;95%CI:2.125-37.776）,CT_Thick（OR,6.531;95%CI:2.410-20.608）,and tumor margin（OR,4.384;95%CT:2.004-9.717）were the predictors of LVI status independent predictors.The ROC analysis showed an AUC value of 0.820（95%CI:0.754-0.885）for the combined model built using multivariable logistic regression.Conclusions:1.Tumor imaging features based on arterial phase enhanced CT can predict LVI status in ESCC patients preoperatively.2.Tumors with LVI are more likely to have homogeneous enhancement patterns,small intratumor vessels,tumor necrosis and irregular tumor margins.3.CT_Tumor,CT_Ratio,CT_Difference,enhancement pattern,intratumor thickened small vessels and tumour necrosis,CT_Thick,CT_Length,CT_Volume,tumor margin morphology correlated with LVI status.4.CT_Ratio,CT_Thickand tumor margin morphology were independent predictors of LVI status.