Noncontrast Computer Tomography-based Radiomics Predictive Study Of Short-term Risk Of Hematoma Expansion In Spontaneous Intracerebral Hemorrhage

Objective: To establish radiomics model based on noncontrast CT images of perihematomal edema and compare it with clinical model,imaging model and combined model,to predict the risk of short-term hematoma expansion in patients with spontaneous intracerebral hemorrhage.Materials and methods: We retrospectively analyzed 139 clinically confirmed patients with spontaneous intracerebral hemorrhage(ICH)in our hospital from January 2014 to January 2021,and relevant clinical data were collected.All patients underwent their first noncontrast head CT scan(NCCT)within 12 hours of symptom onset and repeat CT within72 hours after admission.We defined hematoma expansion as a 33% increase in hematoma volume or >6 m L at the time of the follow-up CT scan according to previous definitions.LIFEx software was applied to delineate the region of interest of the perihematomal edema(PHE)and extract texture features.Then patients were assigned to the training and test groups in a 7:3 ratio according to the principle of randomization.The Z-score algorithm was used to normalize the data.The pearson correlation coefficient was used for dimensionality reduction;ANOVA,RFE,and Relief algorithms were used for feature screening.SVM,random forest,logistic regression,and decision tree algorithms were used as classifiers to construct the radiomics model.Two radiologists jointly assessed NCCT signs.Logistic regression analysis was used to identify independent risk factors for HE.The predictive performance of clinical model,imaging model,radiomics model and the combined model of the above three was compared.Evaluated the advantages and disadvantages of the models by the receiver operating characteristic(ROC)and evaluated the differences between the models using the Delong test.Results:Short-term hematoma expansion occurred in 51 of 139 patients(36.69%).The differences between the two groups of diabetes,PLT,APTT,hematoma edge irregularity,swirl sign and blend sign were statistically significant(P<0.05).In the clinical model,diabetes(OR=3.645;95%CI:1.132-11.735;P=0.030),PLT(OR=0.993;95%CI:0.987-0.998;P=0.013)were independent predictors of hematoma expansion,and the AUCs of training group and testing group were 0.673 and 0.781 respectively.In the imaging model,blend sign(OR=2.983;95% CI: 1.186-7.279;P=0.020),swirl sign(OR=5.380;95% CI:2.121-13.646;P<0.001)were independent predictors of hematoma expansion,and the AUCs of training group and testing group were 0.698、0.763 respectively.The radiomics model established by logistic regression performed best,the AUC value of the training group was 0.845,95% CI: 0.765-0.924,sensitivity: 0.722 and specificity: 0.855;the testing group AUC value was 0.787,95% CI: 0.641-0.933,sensitivity: 0.733,specificity: 0.769.The combined model exhibited higher AUC compared to clinical model,imaging model and radiomics model,and the AUCs of training group and testing group were 0.899、0.944.Conclusion:Diabetes,PLT,blend sign,swirl sign and radscore are independent predictors of hematoma expansion in the short-term.NCCT-based radiomics model can effectively predict short-term hematoma expansion in ICH patients;Combined model constructed by clinical model,imaging model and radiomics model performs better than radiomics model alone.