Content area
Abstract
In this paper, we propose and implement a two-part system for the purpose of automated analysis of intracranial hemorrhage (ICH) in stroke patients using non-contrast computed tomography (CT) scans. The first stage makes use of a convolutional neural network (CNN), specifically a no new U-Net (nnU-Net), to segment hemorrhagic regions in the brain. From the segmented bleeds, radiomic features will be extracted, specifically bleed volume, location, and shape. The extracted features, along with expert-labeled probability scores for the five hemorrhagic subtypes, will be used to train a random forest multi-output regression model. This model will predict the likelihood of each hemorrhage subtype, which includes intraparenchymal hemorrhage (IPH), intraventricular hemorrhage (IVH), subdural hemorrhage (SDH), epidural hemorrhage (EDH), and subarachnoid hemorrhage (SAH). This method takes on a ”glass-box” approach by basing predictions on interpretable features, allowing for clinicians to better understand why certain predictions are made. By making use of the Seg-CQ500 and its corresponding dataset, the CQ500, this system aims to provide further clarity for radiologists by delivering interpretable, feature-driven predictions.