Segmentation Of CT Image Cerebrospinal Fluid Based On Deep Learning

Segmentation of the patient’s cerebrospinal fluid is the basis for the development of a computer-aided brain pathology diagnosis system.Accurately segmenting the area of cerebrospinal fluid in the brain image can be used to evaluate the severity of stroke.In addition,the volume calculated from the segmented cerebrospinal fluid area can also be used to measure the severity of cerebral edema,cerebral atrophy,and hydrocephalus.Therefore,the segmentation of cerebrospinal fluid in medical images has great research value.There is an urgent clinical need for cerebrospinal fluid segmentation with rapid imaging of the brain.First,perform data preprocessing on the original brain CT image: for the fixed distribution of the HU value of the cerebrospinal fluid in the CT image and the unclear boundary with the adjacent tissues,the brain CT image is truncated and normalized,and according to The cross-section CT images are layered to save the pre-processing results.Secondly,the U-Net segmentation method of CT image cerebrospinal fluid is studied.It is found that the U-Net network is directly used to learn the characteristics of CSF,the receptive field of the bottom pixel is 148,and it is impossible to better extract all CSF characteristics.In order to avoid more maximum pooling operations caused by increasing the network depth,U-Net has been improved in the following three ways: First,The jump connection of the fourth layer is changed to a non-Local module,and the equivalent feature map is obtained.secondly Add hybrid expansion convolution at the bottom layer to expand the receptive field of the bottom layer pixels,and better take care of the characteristic information of the details of the cerebrospinal fluid;finally Change the initial channel number of U-Net to 16,reduce Network parameters,reduce memory consumption,and improve the training speed of the model.The network performs trifold cross-validation on the preprocessed data set stored locally.Experimental results show that the U-Net16-nonlocal network,which uses only non-Local modules to expand the receptive field,achieves the highest DSC in Fold3,reaching 0.8521.Its model prediction speed is 0.0188 seconds/layer.But in the Fold2 data set,its Dice is lower than the original U-Net16.In the U-Net16-change network with non-Local module and hybrid expansion convolution at the same time,the cross-validation results of the three are 0.8371,0.8394,and 0.8511,respectively,and the model is more robust.