Research And Implementation Of Medical Image Segmentation Based On Deep Neural Network

With the development of artificial intelligence,how to use deep learning to assist doctors in analyzing medical image data has gradually become a popular research direction.In recent years,there are many models that use deep neural networks to solve medical image segmentation tasks,but these models generally ignore the problem of loss of spatial details caused by multiple downsampling operations in deep neural networks,and usually the spatial details contain a large number of small blocks and edges,etc.The loss of these spatial details will cause overall performance decline of the model,and it will also reduce the medical assistance value of the segmentation result.In response to the above problems,this thesis proposes a medical image segmentation algorithm based on deep neural networks.The main research results are summarized as follows:(1)Aiming at the problem of spatial information loss in existing deep neural network models,this thesis proposes a medical image segmentation algorithm based on spatial information restoration.The algorithm fixes the missing spatial information through the spatial information attention branch(SDAB).One side of the SDAB is directly connected to the shallow features in the encoder,and the advantages of the large-size convolution kernel are used to fully extract spatial information;the other side is connected to the up-sampling features in the decoder to optimize the up-sampling results.SDAB combines attention and feature fusion mechanism to achieve the goal of strengthening spatial information recovery.The experimental results on the LUNA dataset show that SDAB can optimize the extraction and representation of spatial information in deep neural networks,reduce the interference of shallow features on strong semantic information in deep features in the feature fusion stage,and optimize the model’s segmentation of edges and small blocks effect.(2)Aiming at the difficult problem of edge pixel segmentation in medical image segmentation tasks,this thesis proposes a medical image segmentation algorithm based on edge information,which improves the segmentation accuracy of the target edge through the feature enhancement module(FEM).FEM is composed of a channel strengthening block and a pre-guidance block.The channel enhancement block uses the attention mechanism to selectively activate or inhibit channels to enhance the model’s ability to encode and represent semantic information.The pre-guidance block assists the channel strengthening block,uses the auxiliary loss function to generate feedback,improves the channel strengthening block’s ability to select channels,and finally strengthens the deep neural network’s ability to discriminate the target.The experimental results prove that the channel enhancement block and the pre-guidance block can synergistically improve the segmentation performance of the entire model,and the deep neural network model based on the FEM design has a better segmentation performance for edge pixels.(3)Aiming at the difficulty of segmentation of small-scale objects in medical images,this thesis proposes a medical image segmentation algorithm based on deep neural network named spatial detail recovery network(SDRNet).SDRNet uses a modified Resnet encoder,cleverly combines the advantages of SDAB and FEM in the decoder,and uses spatial detail recovery and feature enhancement mechanisms to improve the positioning accuracy of the deep neural network model for small-scale targets.The results of ablation experiments on the LUNA dataset prove that the modules in SDRNet can work together to better handle edge details and small blocks.The comparison experiment with the classic model proves that SDRNet can achieve better segmentation performance under the synergy of the two modules.