Research On Brain Tissue Segmentation Algorithm Based On Neural Network

Human brain tissue structure is extremely complex and can be divided into several regions with specific functions.With the progress of medical imaging equipment and researchers’ continuous research on brain structure,brain segmentation has been divided into three levels: whole brain tissue segmentation,brain structure segmentation and pathological tissue segmentation and extraction.The division of brain tissue into white matter,gray matter and cerebrospinal fluid based on Magnetic Resonance Imaging(MRI)is a long-term hotspot in research,which is related to the problems of disease diagnosis and the understanding of neurodegenerative diseases.However,in practice,the performance of current segmentation algorithms has been seriously degraded due to poor image quality,low contrast and fuzzy boundary caused by the fact that brain MRI is susceptible to noise,migration field,partial volumetric effect and other factors.Therefore,the scientific method of brain tissue segmentation is still a key research topic in the field of medical image processing.This thesis presents two segmentation methods based on neural network.The first is to re-express the MRI image into multi-transform space,and then input the re-expression results into a shallow neural network for training,and finally make prediction by the trained network.The second method proposes a new 3d image segmentation framework for MRI of human brain tissue which is a kind of semantic segmentation method based on deep learning.This framework consists of two parts: the fully convolutional neural network and the long-short-term memory network.The former performs a preliminary segmentation,and the latter further optimizes the segmentation results of the former.The main work of this study as follows:(1)Multiple transformations of human brain MRI images.The Entropy transformation,Laplace transformation and Gabor transformation were performed on each image,and combined with the original grayscale value,we can get a 14 spatial re-expression results for each image.Together with the spatial coordinates,each pixel can express its features as 16 dimensions.(2)Preliminary segmentation using 3d fully convolutional neural network.To build a three-dimensional Full Convolution Neural Network for segmenting the image in pixel level.(3)Refine semantic annotation using a bidirectional long-short term memory network.The grid format data is reconstructed into a sequence format,and then input into a bidirectional long-short-term memory network,and the context knowledge with hidden state is flexibly encoded,thereby enhancing local prediction and solving the weak boundary and defect boundary problem of the MRI image.(4)Results comparison.By comparing the segmentation results of the two methods with the results of other methods,the validity of the three-dimension image segmentation framework is verified.