Segmentation of intracerebral hemorrhage based on improved U-Net

Abstract

Automatic medical image segmentation helps to diagnose and treat stroke timely. In this paper, it is proposing an improved U-Net neural network for the auxiliary diagnosis of intracerebral hemorrhage, which realizes the automatic segmentation of the hemorrhage on CT images. First, clustering the pixels of brain CT images into four categories: white matter, gray matter, cerebrospinal fluid, and hemorrhage by fuzzy C-means clustering method, then removing the skull by morphological image method, and finally proposing an improved U-Net neural network model to segment hemorrhage automatically. Experiments show that the dice similarity coefficient reaches 0.860 ± 0.031, which is better than the other methods. It dramatically improves the accuracy of segmentation for intracerebral hemorrhage.