Research On Cerebrospinal Fluid Segmentation Algorithm For Brain CT Images Based On Improved U-Net Network

The brain is a vital organ of the human body.Early diagnosis and treatment can effectively reduce the damage caused by brain diseases.Cerebrospinal fluid(CSF)segmentation is an important research topic in brain tissue segmentation,which is a key step for brain quantitative analysis.In recent years,CSF segmentation based on deep learning has made great progress,but due to the long prediction time,its extensive application and development is limited.Therefore,it is of great significance to study the segmentation model with high accuracy and high prediction efficiency.In this paper,we propose an automatic CSF segmentation method based on U-Net,namely Reduce U-Net,to improve the efficiency of brain disease detection.The Reduce U-Net is obtained by appropriately modifying the depth and width of U-Net,which can improve the accuracy and speed of the segmentation algorithm.Specifically,compared with the U-Net structure,the depth of the Reduce U-Net is increased from 4 to 5;the width of the network is reduced with the numbers of feature maps in the left branch along the way(16,32,48,64,80),different from that(64,128,256,512)in U-Net.In addition,in order to prevent overfitting of the Reduce U-Net,a dropout layer with the drop rate of0.5 is added to the end of each downsampling layer.It can be seen that the new network reduces the number of training parameters and speeds up the training process.In order to improve the segmentation accuracy of the model when the data set is not large,data augmentation scheme is used to increase the generalization of the model.By analyzing data characteristics of three-dimensional brain CT images,the data were manually labeled using the ITK-Snap software and Materialise Mimics software under the guidance of radiologists.Data augmentation can improve the accuracy of CSF segmentation,without increasing the amount of training samples.The network is trained with a 3-fold cross-validation method on the marked 594 brain CT data.The average dice of the model is 0.8845,and the average volume processing time for thin-slice images is 4.7233 seconds and 0.5513 seconds for thick-slice images.Ablation and comparasion experiments also show that the average prediction time of Reduce U-Net model is greatly shortened under the condition of maintaining a high average segmentation accuracy Dice.