The Prediction Study Of CT-based Radiomics Models And Conventional Radiological Signs On Hematoma Expansion In Acute Spontaneous Intracerebral Hemorrhage

Objective:Spontaneous intracerebral hemorrhage(SICH)is a common subtype of stroke with high morbidity rate and high mortality rate,it is the only factor that could be intervened by physicians after onset.Previous researchers have revealed that conventional radiological signs such as CT angiography(CTA)spot sign can predict hematoma expansion(HE).Radiomics has been widely used in oncology,and already shows preliminary value in intracerebral hemorrhage diseases.Thus,this study explored the value of radiomics models based on non-contrast CT(NCCT)images and CTA images in prediction of spontaneous intracerebral hemorrhage expansion and compared with conventional radiological models.Methods:Retrospectively analyzed the clinical and CT imaging data of 182 patients with spontaneous intracerebral hemorrhage in Northern Jiangsu People’s Hospital from December 2015 to December 2020.All patients underwent the baseline cranial NCCT and CTA scans within 6 hours of onset,and the follow-up CT scan within 48 hours.According to the results of follow-up CT,they were divided into HE group(67 cases)and non-HE group(115 cases),and were randomly divided into training set 145 cases and test set 37 cases using stratified sampling at a ratio of 4:1.At the same time,record relevant clinical data of the patients,including gender,age,blood pressure,Glasgow coma scale(GCS)score,anticoagulant and antiplatelet drugs usage.The conventional radiological signs of the original CT images of the patients,including location,intraventricular hemorrhage,initial volume,shape,swirl sign,blend sign,black hole sign,island sign and CTA spot sign were analyzed,then binary logistic regression analysis models were established.The three-dimensional volume of interest(VOI)along the boundary of the hematoma were delineate on the baseline NCCT and reformated CTA images respectively,1688 radiomics features were extracted.F-test statistic and least absolute shrinkage and selection operator(LASSO)logistic regression was used for feature selection,for unbiased comparison,we empirically set the same number of features as 10.The selected features were combined with logistic regression(LR)machine learning classifier to construct predictive models.Receiver operating characteristic(ROC)curves were drawn and the area under the curve(AUC),sensitivity,specificity and accuracy of the two radiomics models and conventional radiological models were compared to evaluate the effectiveness of each model in predicting HE.Results:The 10 feature parameters selected from NCCT and CTA images were combined with LR classifier to construct radiomics models.In the training set,the AUC,sensitivity,specificity and accuracy of the NCCT model were 0.938,0.849,0.924 and 0.897,respectively.The AUC,sensitivity,specificity and accuracy of the CTA model were 0.904,0.774,0.902 and 0.885,respectively.In the test set,the AUC,sensitivity,specificity and accuracy of the NCCT model were 0.925,0.786,0.913 and 0.865,respectively.The AUC,sensitivity,specificity and accuracy of the CTA model were 0.873,0.714,0.913 and 0.838,respectively.Both showed good predictive performance of HE,and there was no statistical difference between the two radiomics models(P > 0.05).Univariable logistic regression analysis of initial hematoma volume,blend sign and CTA spot sign,showed that they were related to HE in the training set and test set(P < 0.05).The three signs were incorporated into multivariable analysis and the conventional radiological combined model was constructed.The AUC,sensitivity,specificity and accuracy of the combined model in the training set were 0.849,0.811,0.783 and 0.793.The AUC,sensitivity,specificity and accuracy of the combined model in the test set were 0.854,0.857,0.783 and 0.811,respectively.In the training set,the AUC and accuracy of the NCCT model were significantly higher than the three univariable models of spot sign,initial volume and blend sign(P < 0.001),and higher than the conventional radiological combined model(P = 0.022).The AUC and accuracy of the CTA model are also significantly higher than the three univariable models(P < 0.001).In the test set,the AUC and accuracy of the NCCT model were higher than the three univariable models of spot sign,initial volume and blend sign(P < 0.05).The AUC and accuracy of the CTA model were higher than the blend sign model(P = 0.017).Conclusion:Both the radiomics models based on NCCT and CTA images are effective to predict hematoma expansion in patients with SICH,the NCCT model performs best and it could help reduce unnecessary CTA scans,and outperform the predictive effect of conventional radiological models to a certain extent,so as to assist physicians to identify high-risk patients with HE,actively target intervention treatment.