Efficient Brain Hemorrhage Detection on 3D CT Scans with Deep Neural Network

Abstract

A brain hemorrhage is a type of stroke that can cause brain damage and can be life-threatening. The outcome of a brain bleed depends on its location inside the skull, the size of the bleed, and the duration between bleeding and treatment. Brain damage can be severe and results in physical and mental disability. Therefore, saving the lives of such patients completely depends on detecting the correct location of the hemorrhage in an early stage. U-Net is an architecture developed for fast and precise segmentation of biomedical images. Its success in medical image segmentation has been attracting much attention from researchers. In this paper, we propose a novel method for automatic brain hemorrhage detection on 3D CT images using U-Net with a transfer learning approach. The 3D CT images are preprocessed by slicing NIfTI files to 2D, splitting, filtering, and normalization to create input data for our model. We refine and pre-train the U-Net model to detect brain hemorrhage regions on the CT scans. Our proposed method is evaluated on a set of 3D CT-scan images and obtains an accuracy of 92.5%.