Research On Segmentation And Extraction Algorithms Of MRI Spine Images

The human spine is a highly organized and complex part of the central nervous system,which transmits nerve signals between the brain and the peripheral nervous system.Motor neuron degeneration caused by neurodegenerative diseases may cause spinal injury,affect spinal function,and cause various spinal diseases.The severity of spinal diseases and the prognosis of functional recovery are highly dependent on the location and degree of tissue damage.Therefore,there is a clinical need for a spinal segmentation method that can evaluate and characterize the degree of damage of such microstructures on magnetic resonance images,and assist doctors in the diagnosis of spinal diseases.Most of the existing segmentation methods for the spine are relatively simple,with insufficient segmentation accuracy and details,and it is difficult to finely segment the internal tissue of the spinal cord.In response to the above problems,this thesis adopts a two-stage segmentation method,which divides the entire spine segmentation process into two stages: coarse spine positioning and spinal fine segmentation.First,according to the morphological characteristics of the spine,the traditional image processing method is used to perform coarse spine positioning,and then the machine learning method is used to finely segment the spine image after the coarse positioning.In the coarse positioning stage of the spine,this thesis first used the Sobel operator and exponential enhancement to enhance the MR spine image;then,used the level set algorithm to extract the edge;finally,according to the morphological characteristics of the continuous elipse of the biological spinal cord,the Hough transform is used Determine the spinal cord area in the image and complete the coarse positioning process for the spine image.In the spinal cord fine segmentation stage,this thesis proposes a DPD P-net to complete the spinal cord fine segmentation process.First,according to the characteristics of dilated convolution and standard convolution,the model constructs a double-path convolution module to extract features of different receptive fields in the image.Secondly,the pyramid structure is introduced into the double-path convolution module to reduce the semantic gap between encoding features and decoding features.Then,dense skip connections are used for feature fusion inside the pyramid structure to enrich the fine-grained and high-resolution continuous detail feature information in the decoder.In addition,a double-path fusion module is added between the two pyramid structures to fuse the different receptive field characteristics obtained by the two pyramids.Finally,the entire model is trained by depth supervision,so that the model can select a suitable network scale for segmentation prediction when facing segmentation tasks of different complexity.The verification experiment results show that the model proposed in this thesis has good segmentation performance under multiple evaluation indicators on the beagle data set and human data set.Among them,it has excellent performance in three segmentation evaluation indexes of DSC,IOU,and PPV.The visualization results show that the model proposed in this thesis has a significant improvement in the segmentation effect of the complex structure of the gray matter region of the spinal cord.