CT-based Radiomics Study For Diagnosing Gastric Neuroendocrine Tumors

Part One CT-based radiomics study for differentiating gastric neuroendocrine carcinoma from gastric adenocarcinomaObjective: To investigate the feasibility and the clinical value of CT-based radiomics in differentiating gastric neuroendocrine carcinoma(NEC)from gastric adenocarcinoma(ADC).Materials and Methods: From August 2011 to November 2018,the clinical,pathological and imaging data of 48 NEC and 55 ADC patients confirmed by pathology and underwent preoperative non-enhanced and enhanced scanning were analyzed retrospectively.The subjective CT findings including location,shape,confine,cystic change and necrosis,lymph node metastasis and hepatic metastasis were evaluated independently.The thickness,the longest diameter and the CT values of primary lesions in triple-phase were measured and contrast enhancement ratio of every phase were calculated.Meanwhile,the clinicopathological features were recorded.Intraclass correlation coefficient(ICC)and kappa statistics was used to evaluate interobserver agreements for the assessments and measurements.The Student’s t-test(or Mann-Whitney U test)and the Chi-square test(or Fisher’s exact test)were applied to compare the differences of quantitative data and qualitative data,respectively.Logistic regression analysis was used to construct a clinical imaging model.In addition,manual segmentation of lesions with ITK-SNAP software and radiomics features were extracted with Pyradiomics.Extreme Gradient Boosting(XGBoost)was adopted in the process of CT features or radiomics features screening and model building,and the optimal parameters of model was obtained by 10-fold cross-validation.The receiver operator characteristics(ROC)analysis to determine the area under the curve(AUC),accuracy,sensitivity and specificity of the CT imaging model and radiomics model.Delong test was be used to access the differences between the two models.Results: The thickness,confine of tumor,and contrast enhancement ratio of arterial-venous phase were independent indicators for differentiating NEC and ADC group(P<0.05).The three most important important factors in clinical imaging model were contrast enhancement ratio of arterial-venous phase,thickness and confine of tumor,and the corresponding importance scores were 190,174 and 163,respectively.Additionally,the three most important important factors in radiomics model were P_waveletHLL_firstorder_Skewness,PS_original_shape_Surface Volume Ratio and A_original_shape_Flatness,and the corresponding importance scores were 27,21 and 21,respectively.ROC curve analysis showed that the AUC,accuracy,sensitivity and specificity of clinical imaging model and radiomics model were 0.608 & 0.742,58.06% & 71.88%,56.25% & 72.22%,60.00% & 71.43%,respectively.Delong test showed there were significant difference in AUC between these two diagnostic prediction models(P<0.05).Conclusion: 1.The CT findings including contrast enhancement ratio of arterial-venous phase,thickness and confine of tumor can be regarded as valuable factors for differentiating NEC and ADC.2.The CT-based radiomics model is superior to the clinical imaging model in differentiating NEC and ADC,which is of great values to assist clinical quantitative diagnosis.Part Two CT-based radiomics study for evaluating pathological grade of gastric neuroendocrine neoplasmsObjective: To investigate the diagnostic values of CT-based radiomics to determine the pathological grade of gastric neuroendocrine neoplasms(NEN).Materials and Methods: A total of 103 NEN patients confirmed by pathology from August 2011 to November 2018 were retrospectively analyzed,with complete clinicopathological data and preoperative dual-phases enhanced CT scanning of abdominal in our hospital.The subjective CT findings including the number of lesion,location,shape,and lymph node metastasis were evaluated.The thickness,the longest diameter and dual-phases CT values of primary lesions were measured.Meanwhile,the clinicopathological features were recorded.Intraclass correlation coefficient(ICC)and kappa statistics was used to evaluate interobserver agreements for the assessments and measurements.The Student’s t-test(or Mann-Whitney U test)and the Chi-square test(or Fisher’s exact test) were applied to compare the differences of quantitative data and qualitative data,respectively.In addition,manual segmentation of lesions with ITK-SNAP software and radiomics features were extracted with Pyradiomics.Extreme Gradient Boosting(XGBoost)was adopted in the process of CT features or dual-phases radiomics features screening and model building,and the optimal parameters of model was obtained by 10-fold cross-validation.The combined diagnosis model,was built with XGBoost regression analysis using the optimal CT features and radiomics features from CT imaging model and radiomics model.The diagnostic performance of the CT imaging model,the radiomics model and combined diagnosis model for differentiating grade 1/2 and grade 3 NEN was accessed by the accuracy,mean squared error(MSE)and mean absolute error(MAE).The differences in MSE and MAE among three models were accessed by Kruskal-Wallis and pairwise comparisons used by Bonferroni test.Results: The CT findings(the number of lesion,location,shape,lymph node metastasis,thickness and the longest diameter of tumors)and clinical features(age,gender and smoking status of patients)show significant differences between grades1/2 and grade3 NEN(P<0.05).The important factors more than 1400 scores in CT imaging model were the longest diameter of tumors,thickness,CTPP,location,age and lymph node metastasis of tumors.There were six and seven important factors more than 1000 scores in arterial phase and venous phase radiomics model,respectively.The six most important factors in combined diagnosis model were A_logarithm_glcm_Imc1,P_squareroot_glcm_Maximum Probability,thickness,the longest diameter,A_waveletHHL_glrlm_Gray Level Non Uniformity,P_wavelet-LLL_ngtdm_Contrast.For the CT imaging model,the radiomics model and combined diagnosis model,the accuracy were 81.8%,86.0%,87.8% and 91.0%,respectively;the MSE were 539.407,490.076,429.994 and 371.921,respectively;the MAE were 16.716,15.253,14.231 and 12.326,respectively.The combined diagnosis model achived a higest accuracy,lowest MSE value and MAE value,and the performance is higger than that of CT imaging model(P=0.03,P<0.001).Conclusion: 1.Some subjective findings,measurements and clinical features in CT imaging diagnosis could helpful to evaluate and distinguish the grade1/2 and grade3 gastric NEN.2.The CT imaging model,radiomics model and combined diagnosis model was built to differentiate and predict grade 1/2 and grade 3 gastric NEN,and the diagnostic performance of combined diagnosis model is the best among in these models.