CNN Based Hierarchical Intracerebral Hematoma Region Extraction Method in Head Thick-Slice CT Images

Abstract

Cerebrovascular disease is the fourth leading cause of death in Japan, with approximately 100,000 deaths in 2019. The intracerebral hematoma (ICH) is difficult and time-consuming to interpret even for specialists, so an automated method for extracting ICH regions from brain computed-tomography (CT) images is needed to reduce the burden on physicians and improve the speed and accuracy of diagnosis. Because the ICH shows high-absorption in CT images, most of conventional methods segment the high-absorption regions. However, the high-absorption region includes hemorrhagic region such as subarachnoid hemorrhage and intraventricular hemorrhage. These are not ICH. In this study, we propose an automatic extraction method for ICH regions from brain CT images, and the proposed method aims to reduce the over-extraction of high-absorption regions. The proposed method focuses on the anatomical structure of ICH, extracts the high-absorption regions, and proposes a hierarchical method based on classification using convolutional neural network (CNN). The model was trained and evaluated on 33 subjects, and over-extraction was reduced by 82% (specificity) compared to the test data. However, the recall was 36%, and further improvement is needed.