Research On Brain MR Image Segmentation Algorithm Based On SLIC And Deep Learning

With the development of medical imaging technology,Magnetic Resonance Image(MRI)technology can obtain clear images of the internal tissue structure of the human body.Image segmentation technology can help doctors quickly judge the condition and make reasonable diagnosis and treatment plans through computer,so as to save the lives of patients.Accurate segmentation of brain tumors from brain MR Images is a research topic of great significance and challenge.The complexity of brain structure and the characteristics of MR Images limit the accuracy of segmentation algorithms to some extent.Inspired by the concept of superpixels,this paper focuses on brain tumor segmentation based on superpixels and deep learning methods.The specific content is summarized as follows:(1)AG-SLIC(Adaptive Gabor Simple Linear Iterative Clustering)based on adaptive texture perception for superpixel segmentationSuperpixels have a more advanced representation than pixels and are able to reduce noise effects while making the contrast between different regions more obvious.At present,the widely used Simple Linear Iterative Clustering(SLIC)algorithm needs to set segmentation parameters manually and lacks texture perception ability.To solve these problems,this paper proposes an adaptive parameter selection method based on image complexity calculation,and designs Gabor filters of different sizes and directions to enhance the perception of texture features.The experimental results show that the AG-SLIC algorithm can adaptively segment the appropriate number of superpixels,and achieve a good balance between the edge fitting,compactness and regularity of superpixels.(2)Segmentation Network based on Multi-scale information Fusion(MSF-Net)In order to obtain accurate tumor contours from brain MR Images,aiming at the limitations of the existing convolutional neural network model parameters and segmentation accuracy,this paper starts from the expansion of the context information fusion pathway,and designs and constructs a multi-scale information fusion segmentation network with cascading U-shaped network as the backbone.The atrous convolution module with residual Inception structure and the multi-scale pooling module are introduced to increase the acquisition of multi-scale information.The weight sharing module is used as the base layer of the backbone network,so that the number of parameters of the model is kept at a low level.Experimental results show that MSF-Net has higher segmentation accuracy and lower parameter level than the comparison algorithms.(3)Brain tumor segmentation framework combining superpixel and convolutional neural networkAiming at the characteristics and difficulties of brain tumor segmentation task,this paper integrates AG-SLIC algorithm and MSF-Net segmentation network to form a complete brain tumor segmentation framework.Firstly,the input image is processed by median filtering,then the AG-SLIC algorithm is used to mesh the superpixels,and then the enhanced image is obtained by pixel reconstruction operation.Finally,the MSF-Net is used to complete the segmentation of brain tumor contour.Experimental results show that the accuracy of the segmentation framework is improved.