Value Of Intratumoral And Peritumoral Radiomics Based On Contrast-enhanced CT Images In Risk-grade Prediction Of Gastrointestinal Stromal Tumors

Objective: To investigate the value of intratumoral and peritumoral radiomics in predicting gastrointestinal stromal tumors(GISTs)risk-grade using contrast-enhanced computed tomography(CT)images.Methods: Retrospectively enrolled 241 patients with surgically pathologically confirmed primary GISTs in our hospital between January 2017 and December 2021.Risk classification was performed according to the modified NIH criteria,and the enrolled cases were divided into low-risk groups(including very low-risk and low-risk cases)and high-risk groups(including medium-risk and high-risk cases).The included cases were split into the training set and testing set in a ratio of 7:3.The clinical,pathological and traditional CT features of the eligible patients were collected.The characteristics with P<0.1 between-group comparison were screened by univariate analysis and included in the binary logistic regression(LR)analysis to obtain independent risk factors associated with the GISTs risk classification,and the clinical model was constructed by LR.The intratumoral,3mm peritumoral and 5mm peritumoral regions of interest(ROI)were segmented layer by layer from CT venous phase thin-layer images by a radiologist(with 7years of radiology experience)using 3D-slicer software.Then extracted the radiomics features by Python software after pre-processing.Twenty cases were randomly selected to be outlined by another radiologist(with 8 years of radiology experience)for consistency assessment,and the radiomics features with inter-observer correlation coefficient(ICC)greater than 0.75 were retained.Five radiomics models were established by LR after features selection by one-way analysis of variance(ANOVA)and least absolute shrinkage and selection operator(LASSO).The optimal radiomics model was selected to calculate the radiomics score(Rad-score),and the independent predictors of clinical,pathological and traditional CT features were combined with the Rad-score to establish a combined clinical-radiomics model and draw a nomogram.The area under the curve(AUC),sensitivity(SEN),accuracy(ACC),and specificity(SPE)were used to assess the predictive ability of each model.Differences in AUC values between models were compared using Delong’s test.The calibration degree of the clinical,optimal radiomics model and combined clinical-radiomics model was evaluated by the calibration curve.The clinical value of the clinical,optimal radiomics model and combined clinical-radiomics model were compared using decision curve analysis(DCA).Results: 1.The 241 GISTs patients were divided into training set(n=169;including 78 low-risk cases and 91 high-risk cases)and testing set(n=72;including 33 low-risk cases and 39 high-risk cases).The results of univariate analysis and bivariate LR analysis revealed that the maximum diameter of tumor(P<0.001;OR,4.214;95%CI,2.687-6.609),Ki-67 index(P=0.001;OR,6.661;95%CI,2.225-19.943)and tumor necrosis on CT images(P=0.014;OR,0.385;95%CI,0.180-0.823)were independent predictors for the risk classification of GISTs.Based on the factors mentioned above,the AUC was0.913(95%CI: 0.860-0.95),SEN was 0.802,ACC was 0.787 and SPE was0.885 of the clinical model(CM)in the training set.The AUC was 0.874(95%CI: 0.774-0.940),SEN was 0.872,ACC was 0.792 and SPE was 0.758 in the testing set.2.A total of 1200 radiomics features were extracted from each of the intratumoral,3mm peritumoral and 5mm peritumoral ROIs,and most of the features were retained after ICC evaluation.The radiomics features from the3 mm peritumoral region and 5mm peritumoral region were respectively combined with the radiomics features from the intratumoral region.Finally,five radiomics models were established including one intratumoral model(ITV),two peritumoral models(PTV3,PTV5)and two combined intratumoral and peritumoral models(ITV+PTV3,ITV+PTV5).Among them,combined intratumoral and 3mm peritumoral model(ITV+PTV3)showed the best prediction performance of GISTs risk classification.The AUC was 0.970(95%CI:0.932-0.990),SEN was 0.901,ACC was 0.905 and SPE was 0.923 in the training set of ITV+PTV3.The AUC was 0.954(95%CI:0.877-0.989),SEN was 0.846,ACC was 0.889 and SPE was 0.939 in the testing set.Delong’s test revealed that ITV+PTV3 was superior to CM both in the training and testing sets(P<0.05).The calibration curves(CC)showed that ITV+PTV3 have good agreement with the ideal curves,and had the best fitness in the testing set.The decision curves indicated that ITV+PTV3 had a higher net clinical benefit than CM over a wider range of threshold probabilities.3.The Rad-score of ITV+PTV3 was calculated.The combined clinical-radiomics model(CRM)established by combining the maximum diameter of tumor,Ki-67 index,tumor necrosis on CT images and Rad-score showed the highest AUC value among all the models.The AUC was 0.979(95%CI:0.944-0.995),SEN was 0.890,ACC was 0.905 and SPE was 0.974 in the training set.The AUC was 0.960(95%CI:0.885-0.992),SEN was 0.897,ACC was 0.889 and SPE was 0.879 in the testing set.In both the training set and the testing set,the calibration curve showed good calibration and DCA showed high clinical net benefit.Delong’s test showed that CRM was superior to the CM(P<0.05),but the AUCs were not significantly different from those of ITV+PTV3.Conclusion: The radiomics models and combined clinical–radiomics model based on intratumoral and peritumoral regions of venous phase CT images perform well at predicting the risk classification of GISTs,which showed well clinical application value and can effectively assist accurate preoperative diagnosis and treatment decision of GISTs.