Prediction Of Hemorrhagic Transformation In Acute Ischemic Stroke Based On Texture Analysis Of CT Images

Objective: Hemorrhagic transformation(HT)is the most dread complication of cerebral infarction after endovascular therapy,which can affect the prognosis of patients to a large extent.Therefore,many researchers attempted to predict the risk of bleeding after treatment of patients with cerebral infarction via the early stage scanning images or clinical data.This study attempted to do the HT prediction through texture analysis in computed tomography(CTA and NCCT)images,and compared the predictive efficacy with traditional clinical scoring methods.Methods: 141 patients with acute ischemic stroke(AIS)were enrolled in this study including 73 endovascular therapy(EVT)cases and 68 conservative treatment(CST).We divided the research objects into three groups: EVT,CST and EVT+CST mixed group.Each group was further divided into three subgroups: CTA infarct area,CTA contralateral non-infarct area and CT infarcted area.In each group,the ROI of the infarction area was extracted according to the apparent diffusion coefficient(ADC)diagram and matched to each subgroup image.Then,HT related texture features were extracted from the target ROI of each subgroup.Finally top ten important features were selected and substituted into the classifier for training,and the classification performance was evaluated through 5-fold cross-validation.Four clinical predictive scores(HAT,SEDAN,THRIVE-c,and HIAT2)were obtained for patients in the EVT group,and the predictive efficacy of their interaction and texture analysis was compared by ROC curve.Results: The classification model performed well both in EVT and CST group for texture analysis in infarct region of CTA,with a performance of 0.7810±0.0575 accuracy,0.8819±0.0544 AUC,and 0.7400±0.0514 F1 score for EVT group,and0.7670±0.1356 accuracy,0.8478±0.1314 AUC and 0.6732±0.1698 F1 score in CST group.The classification performance of model based on infarct region in CTA was obviously superior to contralateral non-infarct area in CTA and infarct region in noncontrast CT,and the separate EVT and CST groups was obviously better than the mixed group.HAT,SEDAN,THRIVE-c and HIAT2 were significantly correlated(P < 0.001),with AUC values of 0.678,0.685,0.640 and 0.607,respectively.Conclusion:1.Texture analysis based on CT images(CTA and NCCT)has the ability to predict HT,among which infarct area of CTA has the best prediction efficiency,providing valuable information for clinical treatment decisions of ischemic stroke;2.The prediction method of texture analysis is vastly better than the traditional clinical prediction scores of HAT,SEDAN,THRIVE-c and HIAT2.