Research On Automatic Segmentation Method Of Cerebral Hemorrhage In CT Images

Stroke is the second most common cause of death in the world and one of the main causes of disability.CT is considered to be the first choice for diagnosis of acute stroke patients due to its fast imaging speed,low cost,and high availability.The automatic segmentation method of medical images provides great help for the timely diagnosis and treatment of stroke.Aiming at the problem of assisted diagnosis of cerebral hemorrhage,this paper combines traditional segmentation methods and deep learning methods to propose an automatic segmentation method for the hemorrhage area of ? ? stroke CT images.Aiming at the problems of blurred edges of cerebral hemorrhage in brain CT images,small amount of data,and difficulty in manual segmentation,this paper proposes an improved U-Net neural network to automatically segment the hemorrhage area.First of all,because the more network layers,the more complex the architecture,the more parameters,the more time it takes to calculate,and the easier it is to overfit.Therefore,the overall structure is adjusted to seven layers and only three downsampling is used.Secondly,the maximum pool operation will halve the number of features,and adding a convolution operation during the jump connection process can better extract features without increasing the number of network layers,which is convenient for improving the accuracy of segmentation.The experiment uses a variety of data enhancement methods to increase the amount of data,and the results show that its Dice Similarity Coefficient(DSC)is 0.837,which is better than the classic U-Net neural network.When segmenting the hematoma area,the skull is easy to be mistaken for the hemorrhage area,which is the main factor hindering the segmentation of the hematoma.In order to solve the above problems,a combination of fuzzy C-means clustering method and morphological method for preprocessing is proposed.The image is clustered into four parts:gray matter,white matter,cerebrospinal fluid and skull,and then the skull is removed by expansion and erosion.And further optimize the model,add multiple cross-scale splicing operations,that is,each result obtained in the down-sampling process will be spliced ??with the up-sampling process before convolution,to add more information and make the segmentation effect more accurate.The experimental results show that the proposed improved U-Net neural network model is better than the classic U-Net,white matter fuzzy C-means clustering,multi-path context generation confrontation network and other methods.The experiment also compares the results of three different loss functions.It is found that the cross-entropy loss function has the best effect,and the dice similarity coefficient can reach 0.860±0.031.The topic aims at the CT image segmentation of cerebral hemorrhage to assist in the diagnosis of problems,and proposes an improved U-Net network model,which combines multiple models and further optimized U-Net to achieve automatic segmentation of the bleeding area,which greatly improves hemorrhagic stroke The accuracy of region segmentation.The proposed method can be applied to the surgical navigation system.The improvement of the method can improve the accuracy of segmentation of hematoma,reduce the operation time,reduce the burden on the doctor,and provide great help for the treatment of hemorrhagic stroke patients.