Research On Medical MRI Image Segmentation Methods Based On Deep Learning

Image segmentation is an important research content in the field of artificial intelligence and computer vision.Medical image segmentation refers to dividing the medical image into several regions and accurately extracting the regions of interest in the image,which is the key issue to determine whether the medical image can provide a reliable basis in clinical diagnosis and treatment.it is of great significance to clinical applications such as disease diagnosis,surgical planning,and prognosis evaluation.The purpose of this paper is to study the medical image segmentation method based on deep learning,especially to solve the problems of small focus area,insignificant boundary,complex and changeable contour in MRI image,an image segmentation algorithm based on an attention mechanism is proposed,and a good segmentation result is obtained on the well-known data set.The main research contents and contributions of this paper are as follows:(1)To solve the problems of insufficient global information extraction and low training efficiency in the mainstream U-Net-based segmentation model,a segmentation network SAU-Net based on U-Net and the self-attention mechanism is proposed.The model uses the self-attention module to increase the global information and reduce the network parameters.A fast and concise decomposition and convolution method are designed to solve the problem of low training efficiency,and the residual connection is integrated into the network.The experimental results show that the number of parameters of the SAU-Net model is less and the Dice coefficient is better.(2)To solve the problem of insufficient detail feature extraction of small lesions in the U-Net segmentation model and the inefficiency of the decoder using high-and lowlevel features,a deformable attention mechanism,and an optimized segmentation network Def A-Net are proposed.The model adopts the idea of deformable convolution,pays attention to the edge of the focus through the focus bounding box of different scales and angles,and makes full use of the context information to extract the details of the focus.A weighted sampling fusion decoder is designed to make use of the influence of different high-level features on the segmentation results to improve the segmentation accuracy and efficiency.The experimental results show that the method proposed in this paper has better performance in terms of parameter number,robustness,training efficiency,accuracy,and structural similarity.(3)Based on the idea of transfer learning,the proposed method is used to segment and optimize the stroke focus detection MRI dataset Stroke QD established by the research group and the hospital at the pixel level,which verifies the generalization ability and applicability of this method.