Deep learning prediction of contrast extravasation versus intracranial hemorrhage after thrombectomy in patients with acute stroke

It is difficult to distinguish the nature of contrast extravasation (CE) vs. intracranial hemorrhage (ICH) on immediate postprocedural computed tomography (CT) after endovascular thrombectomy. This differentiation between CE and ICH carries important clinical implications, and this study aims to explore and validate the feasibility of approaches based on machine learning algorithm. Patients with the hyperdensity on the immediate post endovascular treatment (EVT) CT scans were enrolled. Clinical and radiologic data of these patients were collected and nature of hyperdensity was identified based on the comparison of the immediate post-EVT CT scans vs. CT scans obtained 24 h post-EVT. After images with hyperdense lesion were labelled, two deep learning models based on convolutional neural network (CNN) were derived and validated. Model 1 was derived and internally evaluated with five folds cross-validation, and model 2 was fitted and evaluated on radiographs of whole patients who were randomly divided into the training set, validation set, and testing set. Moreover, the performance of model 2 on the testing set was compared with a support vector machine (SVM) model and a recursive partitioning and regression trees (RPART) model based on CT Hounsfield Units (HU) values. A total of 106 patients were enrolled, 63 patients were identified as CE, and 43 as ICH. Model 1 accomplished classification performance with the mean AUC of 0.955 ± 0.024 on the validation set. The performance of model 2 reached an AUC of 0.956 on testing set, which was higher than those of SVM model and RPART model. CNN-based deep learning algorithm demonstrated favorable classification performance in distinguishing between CE and ICH on post-EVT brain CT scans, and it provide a feasible exploratory tool for this clinical differentiation task.