基于改进U-Net的CT图像脑出血区域分割方法

SEGMENTATION OF INTRACEREBRAL HEMORRHAGE AREA IN CT IMAGES BASED ON IMPROVED U-NET

医学图像自动分割方法对中风的及时诊断和治疗提供了极大的帮助。实验针对脑出血辅助诊断问题,提出一种改进的U-Net神经网络模型,实现脑中风CT图像出血区域自动分割。对脑部CT图像进行预处理,用模糊C-均值聚类方法将图像聚类成脑灰质、脑白质、脑脊液和出血区域4类;通过形态学图像方法去除颅骨;提出一种改进的U-Net神经网络模型,并用该模型进行出血区域的自动分割。实验结果表明,提出方法的骰子相似系数(Dice Similarity Coefficient, DSC)可达0.860±0.031,优于白质模糊C-均值聚类、多路径上下文生成对抗网络等方法,极大提高出血性中风区域分割的准确率。

Automatic medical image segmentation method greatly helps to timely diagnose and treat stroke.In this paper,an improved U-Net neural network is proposed for the auxiliary diagnosis of cerebral hemorrhage,which realizes automatic segmentation of the hemorrhage area in CT images of stroke.The CT images of the brain were preprocessed and clustered into four categories:gray matter,white matter,cerebrospinal fluid,and hemorrhage regions by fuzzy C-means clustering method.The skull was removed by morphological image method.An improved U-Net neural network model was proposed,which was used to automatically segment hemorrhage areas.Experimental results show that the dice similarity coefficient of the proposed method reaches 0.860±0.031,which is better than white matter fuzzy C-means clustering,multipath context generation adversarial network and other methods.It greatly improves the accuracy of segmentation for hemorrhagic stroke area.