Application Of CT Radiomics Combined With Clinical Features In Predicting Risk Classification Of Gastrointestinal Stromal Tumors

Objectives:Radiomics technology has made progress in the clinical diagnosis and prognosis of gastrointestinal stromal tumors(GIST)due to the advantages of non-invasive and quantitative analysis of mega data.However,there are differences in the phases of CT examination during image acquisition and the readability of the model is weak.The objective of this study was to find the best predictors from clinical data and CT signs,phases-based enhanced CT radiomics and the combined nomogram model respectively and then construct the diagnostic tool for the application of non-invasive and quantitative preoperative assessment of risk classification of GIST.Methods:The study collected patients with postoperative pathological diagnosis of GIST retrospectively,who visited the Second Hospital of Jilin University from January2016 to September 2022 and underwent CT plain plus enhanced examination.A total of 105 patients satisfied the inclusion criteria.The patients were divided into high and low malignant potential groups in a 7:3 ratio based on the results of pathological risk grouping.The grouping results were analyzed for baseline concordance.Firstly,12features related to clinical data and CT signs were collected and recorded.High-risk features were screened using univariate analysis,and then correlation analysis and binary logistic regression analysis were performed to find independent predictor associated with high malignant potential of GIST.Subsequently,3D ROI(Three-dimensional region of interest,3D ROI)was operated on reconstructed 3 mm CT examination four-phase images according to the outline criteria.The best feature subset was obtained after filtering and downscaling features used by A.K.and IPMS softwares,and then CT contrast-enhanced model and CT flat-scan plus contrast-enhanced model were constructed.The diagnostic performances of the models were analyzed in terms of ROC(Receiver operating characteristic,ROC)curve,AUC(Area under the curve,AUC)value,accuracy,sensitivity,specificity,positive predictive value and negative predictive value.The difference of AUC of the models was validated using Delong Test.Finally,nomogram model was conducted by the combination of the optimal features from clinical imaging data and CT radiomics.The performances among models were compared in terms of diagnostic accuracy,calibration,and net clinical benefit to patients.Results:1.The data for all patients in the training and validation groups were consistent at baseline(P>0.05).In the analysis of clinical imaging data,the results of univariate analysis showed age and gender had no statistical differences,and the results of CT signs indicated the existence of five high-risk features including tumor growth pattern,morphological change,size,liquefaction or necrosis,and enhancement pattern(P<0.05).Correlation analysis about high-risk features showed the strong correlation between tumor size,liquefaction or necrosis and malignant potential of GIST(r_s=0.827,0.705),and morphological change and enhancement pattern were moderately correlated(r_s=0.655,0.514).The results of binary logistic regression analysis confirmed that the maximum diameter of the tumor was the independent risk predictor(OR=1.141[95%CI,1.058～1.231],β=0.132).2.Both the CT contrast-enhanced model and the CT plain plus contrast-enhanced model showed excellent diagnostic performance in the prediction of malignant potential of GIST(training group,AUC=0.950 vs 0.976,validation group,AUC=0.906 vs 0.969,Delong Test,P=0.087).Furthermore,the diagnostic performance of the CT plain plus contrast-enhanced model was better than that of the CT contrast-enhanced model overall,especially in accuracy,sensitivity,and negative predictive value(accuracy=0.918 vs 0.877,sensitivity=0.865 vs 0.784,and negative predictive value=0.875 vs 0.814 in the training group,more significant differences in the validation group).The CT flat-scan plus contrast-enhanced model also demonstrated excellent specificity and positive predictive value(training group,0.972and 0.970,validation group,0.938 and 0.933).Two models had good calibration and clinical use,and the net clinical benefits to patients ranged from 0.3 to 0.5.3.Nomogram model was conducted based on the combination of the Radscore of CT plain plus contrast-enhanced model and tumor size.Compared to single CT radiomics model,the nomogram model showed higher diagnostic accuracy(training group,AUC=0.985,validation group,AUC=0.972).The calibration and decision curves showed that the nomogram model had better performance in calibration and clinical use,with a net clinical yield of 0.6 to 1.Conclusions:The maximum diameter of tumor is an independent predictor of the risk classification of GIST.Among the CT radiomics models,the CT flat-scan plus contrast-enhanced model shows superior diagnostic performance.The nomogram model based on the combination of the Radscore and tumor size exhibits higher diagnostic accuracy,calibration,and clinical usability,which is promising as a non-invasive,quantitative tool with high diagnostic value for the application of preoperative assessment of the malignant potential of GIST.What’s more,the nomogram model can be easily operated and holds promise for widespread use in the clinical treatment management of GIST.