Segmentation Of Intracranial Aneurysms Based On Convolutional Neural Network

With the development of the Internet of things information industry,the amount of various kinds of information is growing explosively,especially in the medical field,such as image data,video data and text data,etc.,which has important potential value.With the continuous development of big data analysis technology,there is a deep combination of deep learning and medical field.For example,the detection of intracranial aneurysms by deep learning technology is one of the current research hotspots.Intracranial aneurysm refers to the cystic protrusion of the arterial wall,which is the first cause of subarachnoid hemorrhage.The rupture of the aneurysm is related to its size,and it is slow and tedious to label the aneurysm only by doctors,especially for large datasets.This thesis focuses on the study of non-local neural network,improves different aspects of the network,and applies it to the detection and measurement of intracranial aneurysms.The main work of this thesis is as follows.(1)We propose a new non-local neural network with improved the similarity measure function to solve the problem that the similarity measure function in non-local neural network is likely to cause insufficient weight,Firstly,the similarity between the output position and all possible positions is calculated by Gaussian function,and then the similarity is calculated by cosine function.The two similarities are combined by point multiplication,which makes up for the deficiency of Gauss weighting and cosine over weighting.Secondly,the improved non-local module is inserted into the convolutional neural network,so that the extracted features fully consider the non-local information.The experimental results show that the effectiveness of the proposed neural network is significantly improved.(2)U-net has not achieved the expected results in some medical image segmentation due to the complexity of the size and shape of the segmented object.To solve this problem,an improved non-local neural network based on u-net is proposed.Firstly,using the u-net neural network under sampling process for feature extraction,after the second layer of this process,the non-local module is embedded,so that the extracted features can better grasp the non-local feature information.Secondly,the information of the down-sampling is integrated with the up-sampling process,and the information of each level is retained by "skip connection",so that the parts which need to be segmented in the image can be accurately located.The experimental results show that the u-net neural network combined with non-local modules has a significant improvement in image segmentation.