Research On Medical Image Segmentation Based On U-Net Model

With the continuous development of medical engineering technology,medical image segmentation plays an important role in pathological diagnosis,surgical guidance and postoperative recovery.At present,U-Net model based on deep learning has been widely used in the field of automatic segmentation of medical images.In this paper,automatic segmentation of medical images is realized by combining UNET model and other image processing methods.The main research contents are as follows:(1)To solve the problem that traditional 2D convolution cannot effectively extract spatial information,a new Att-Dial Res Net3 D model based on attention mechanism is proposed.The Att-Dial Res Net3 D model was improved on the basis of U-Net framework.The sawtooth expansion residuals convolution block is used to extract the inter-chip and intra-chip features of the 3D image to reduce the loss of spatial information,and the context information is fused by the method of attention.The proposed model is verified on the Li TS2017 liver tumor segmentation dataset.The experimental results show that the average DICE of Att-Dial Res Net3 D model for liver and tumor segmentation reaches 0.958 and 0.666,respectively,and the global DICE reaches 0.962 and 0.800,respectively.(2)To solve the problem of gradient dispersion and network degradation caused by too many convolutional layers and pooling layers in the original U-Net model,a U-Net model based on pyramid residuals is proposed.In this model,image features of different scales are obtained by using pyramid residuals,and these features are fused with the features extracted from the convolutional layer,so that richer feature information can be obtained without adding network parameters.The proposed model was validated on the ISIC 2018 skin cancer segmentation dataset,and the experimental results show that the U-Net segmentation model based on pyramid residuals has higher accuracy and better robustness.(3)In order to solve the problem that stackable local operations in U-NET model cannot efficiently obtain global information,a multiattention based U-Net model is proposed.The model establishes the dependency among all features through self-attention,and uses channel attention and spatial attention to reduce the information loss in the pooling process.The segmentation model was verified on MICCAI 2020COVID-19 segmentation dataset and SCUI 2020 segmentation dataset for thyroid nodules.The experimental results show that the U-Net model based on multi-attention can achieve global information aggregation and efficient and rapid segmentation.