Siamese U-Net with Healthy Template for Accurate Segmentation of Intracranial Hemorrhage

Intracranial hemorrhage (ICH) is a fatal form of stroke which is caused by bleeding within or around the brain. Detection and quantification of hemorrhage are critical in the diagnosis and treatment of the disease. In this paper, we propose Siamese U-Net, to segment the abnormal regions of ICH more accurately from patients’ CT images. The Siamese U-Net is given a paired set of the patients’ CT images and a healthy template of the brain CT. We introduce the dissimilarity of hemorrhage regions from the healthy template to the long skip-connection in the U-Net architecture to emphasize the convolutional features of the abnormal regions by ICH. We evaluate the accuracy of the proposed architecture with a comparison of the baseline model. The proposed model shows significant improvement in Hausdorff distance (6.81%), dice score (9.07%), and volume percentage error (40.32%), compared to the baseline U-Net model. Regarding the healthy template, less both false-negative and false-positive regions are observed in the results of the Siamese U-Net. Consequently, the estimated blood volume by the Siamese U-Net is much closer to the actual volume than that of the baseline U-Net.