Objective:The purpose of this study was aimed to evaluate the efficacy of a deep learning radiomics(DLR)model based on non-contrast CT scans for predicting early hematoma ex-pansion(HE)in patients with spontaneous intracerebral hemorrhage(sICH).Methods:The compre-hensive clinical and imaging data of 350 sICH patients treated at our institution from Jan 2015 to Dec 2022 was retrospectively analyzed.All patients underwent an initial head CT scan within 6 hours of symptom onset.Patients were divided into HE group(136 cases)and non-HE group(214 cases)based on whether the hematoma volume increased by more than 33％or 6mL according to a follow-up CT performed within 24 hours.Patients were randomly assigned in an 8∶2 ratio to a training group(n=280)and a validation group(n=70).Clinical and imaging features were compared between the two groups,and statistically significant features were identified.Regions of interest(ROIs)were manually delineated layer by layer along the hematoma border,and three-dimensional volume ROIs(VOIs)were generated.Software was used to automatically expand the ROIs by 2mm to include peri-lesion tis-sue.Radiomics features and deep learning features(based on ResNet-50 convolutional neural network)of the peri-lesion tissue were extracted using One-key AI software;and then,the features of the two types were combined and selected to create a set of mixed features.Based on clinical and imaging fea-tures,mixed features and their combination respectively,15 machine learning(ML)models were de-veloped using five classifiers,including logistic regression(LR),naive Bayes(NB),K-nearest neigh-bors(KNN),adaptive boosting(AdaBoost)and multilayer perceptron(MLP).The diagnostic efficacy of each model was assessed by the area under the receiver operating characteristic curve(AUC),and the best-performing model was selected as the final output model.Decision curve analysis(DCA)was conducted to evaluate the clinical utility of the optimal model.Results:In the clinical and conventional imaging features,serum D-dimer level,hematoma shape,swirl sign,mixed density sign and satellite sign showed statistically significant difference between the HE group and non-HE group(all P＜0.05).A total of 29 mixed features(15 radiomics features and 14 deep learning features)were extrac-ted from perilesional tissue.Models based on combined features outperformed those based on clinical-imaging features or mixed features alone in both the training and validation groups.The KNN classifier model based on combined features demonstrated the highest predictive performance in the training group(AUC=0.947,95％CI:0.924～0.970)and was selected as the optimal model.DCA indicated that the KNN combined model provided good clinical benefit across a probability threshold range of 0.025 to 0.980.Conclusion:The DLR model based on CT scans of perihematomal tissue effectively predicts early hematoma expansion in sICH cases.The KNN classifier model constructed by clinical,imaging and radiomics features offers the best predictive performance.
Deep LearningClassifiersRadiomicsPerihematomal tissueSpontaneous in-tracerebral hemorrhageHematoma expansion
万方数据
维普
