Application Research Of Aneurysmal Subarachnoid Hemorrhage Detection Based On UNe

Aneurysmal subarachnoid haemorrhage(a SAH)is the most common nontraumatic subarachnoid hemorrhage(SAH),with high mortality and complication rates.a SAH severely damages the central nervous system and has pathological effects on multiple organs throughout the body.The clinical treatment of a SAH is relatively complicated.Due to the knowledge and technology of multiple disciplines involved,doctors from multiple disciplines are required to form a medical team to treat patients,which requiring a lot of manpower.Such high demand affects the early treatment of a SAH,but early assessment and treatment of the patient’s disease is the key to reducing a SAH mortality and morbidity.In recent years,Artificial Intelligence(AI)technology has developed rapidly,especially in the auxiliary diagnosis of imaging,making it possible to assist doctors in diagnosis through AI technology.This thesis starts with the UNet network,analyzes the analyzes the derived network of UNet,summarizes its improvement ideas,combines the image features of a SAH,and combines the image features of a SAH to propose a semantic segmentation network RBA-UNet for a SAH detection,so that the trained a SAH model can better detect a SAH regions in head Computed Tomography(CT)Scans with improved segmentation performance.Firstly,this thesis starts with the existing research results,analyzes the existing head CT diagnostic services,and proposes a solution to the problem of spatial information loss caused by the continuous convolution span and pooling operation of the UNet structure by analyzing the derived network of UNet.Secondly,the image features of a SAH are analyzed,for the diversity of shape and position of a SAH and the complexity of edges in head CT,Atrous Spatial Pyramid Pooling(ASPP)is used to achieve arbitrary multi-scale feature extraction.Aiming at the problem of spatial information loss caused by the continuous convolution pooling of UNet network,the residual structure is used to improve the decoder part,and Batch Normalization(BN)and Rectified Linear Unit(Re LU)are used to speed up the training and convergence speed of the network,and to solve the problem of gradient explosion and gradient disappearance while preserving spatial information.Finally,for the 3D a SAH data of 3D Slicer,uses python and matlab to generate a 2D dataset of a SAH images,uses python and matlab to create a two-dimensional a SAH data set CA80,uses the learning rate decay and cross-validation strategy,and uses the RBA-UNet network to train the a SAH segmentation model,and conduct experimental analysis with the existing solutions on the market and models trained by other improved UNet algorithms.The test results show that for the dataset CA80,the accuracy of the a SAH segmentation model trained based on the improved UNet network RBA-UNet in this thesis reaches 87.74%.Compared with the existing head CT diagnosis service and UNet improvement scheme,the RBA-UNet proposed in this thesis has better accuracy in the segmentation of a SAH,which basically meets the requirements of medical assistance.