Enhanced Detection of Intracranial Hemorrhage: A New Hybrid Model Design based on the U-Net Segmentation Method

Intracranial hemorrhage (ICH) denotes bleeding inside the skull, which can occur in or around the brain. Computed tomography (CT) has been used to detect ICH due to its high efficiency and accuracy. Nowadays, deep learning model design is introduced to allow an accurate and efficient classification of ICH in CT images. This work focused on developing U-Net-based models for the segmenting of ICH. Furthermore, the proposed model employed two transfer learning models, MobileNet and Xception, as the backbones of the U-Net topology. This approach aims to establish metrics that improve ICH treatment through precise segmentation techniques. A free dataset from the Radiological Society of North America (RSNA) was fed to the proposed model. This dataset comprises 5996 annotated CT images divided into four groups, with 1499 images per group. Three of these groups are for hemorrhage, and one is for the normal group. The average detection accuracy rates were impressive, with Intraparenchymal Hemorrhage (IPH) at 98%, Intraventricular Hemorrhage (IVH) at 98%, and Subdural Hemorrhage (SDH) achieving accuracy of 97%. This approach could assist radiologists in overloaded medical centers in precisely detecting ICH, especially in the local healthcare centers.