Glioma Segmentation Based On Deep Learning And 3D Reconstruction

Glioma is a primary intracranial tumor that originates in the brain's glial cells,it affects human health and quality of life seriously.Improving the accuracy of glioma segmentation helps to improve the accuracy of clinical diagnosis and reduce the rate of misdiagnosis.The diversity and complexity of gliomas makes it difficult for glioma segmentation.Therefore,it's high research value to improve the accuracy of glioma segmentation.However,it is still a of slice images after glioma segmentation.When performing clinical diagnosis,doctors still need to rely on their rich medical knowledge and spatial imagination to reconstruct the characteristics of glioma in maind,which is likely to cause misdiagnosis and missed diagnosis.The sequence of images could be rendered to a three-dimensional image by 3D reconstruction technology,which provides more basis for doctors to conduct clinical diagnosis and formulate treatment plans.Therefore,three-dimensional reconstruction of medical images has important application value.In view of the relatively few features of single-path feature extraction,resulting in the problem of low segmentation accuracy of the target area,this paper uses the method of dual-path feature extraction to extract richer feature information for improving the segmentation accuracy.Because the feature fusion caused by the operation of concatenate feature maps in the U-net network is insufficient and the large amount of calculation,this paper uses a 1×1 bottleneck layer to fuse the feature maps in the decoding process more fully,while reducing the amount of parameters effectively.In view of the problem about slow network training speed caused by the change of mathematical distribution of feature vectors in the process of network training,and the problem may appear to be overfitting,the Batch-Normalization layer is followed the feature vectors to speed up the network convergence speed to effectively prevent the occurrence of overfitting.There are a lot of invalid voxels in the 3D reconstruction process,which will generate too many meaningless interpolation operations and cause the 3D reconstruction time to be too long.For this,whether the sampling point is invalid is judged,and only interpolates the validsampling point,thus improve reconstruction efficiency.In view of the problem of poor visual effects at the edges of different areas in 3D reconstruction,this paper judges the category attributes of each sampling point firstly,assigns different color attributes to the sampling points with different category attributes,and improves the accuracy of edge judgement about the areas in the 3D reconstruction image.Finally,the feasibility and effectiveness of the image segmentation algorithm and 3D reconstruction algorithm used in this paper are verified through experiments.