YOLOv3-based Intracranial Hemorrhage Localization from CT Images

Abstract

Intracranial hemorrhage (ICH) is a common stroke type that requires an early and urgent diagnosis. The standard imaging modality for ICH diagnosis is computed tomography (CT). However, the type of hemorrhage must be identified by the neurologist to make an effective treatment decision. Although available traditional methods and deep learning-based algorithms for ICH detection can achieve excellent performance, the classification and segmentation of ICH images are difficult tasks since multiple types of ICH may exist within the CT image. Localizing ICH through a bounding box is a more straightforward task than the semantic segmentation task where the model tries to classify pixel-wise. In this work, the YOLOv3 model is proposed to localize mixed hemorrhages from CT images. Additionally, a pipeline for data augmentation was applied to address the problem of limited bounding box annotations for ICH detection. The YOLOv3 model has been evaluated and validated on the brain hemorrhage extended dataset. The proposed method achieved competitive results against state-of-the-art methods.