Research And Implementation On Key Technologies Of Medical Image Segmentation Based On Encoder-decoder Architecture

Medical image segmentation plays an important role in clinical diagnosis,quantitative analysis and treatment process.Segmentation model based on encoderdecoder architecture has been widely used in medical image segmentation.In actual segmentation,improper design of encoder,skip connection and decoder components in encoder-decoder architecture leads to improper multi-scale feature fusion,similar features are not related,semantic gap caused by direct feature channel concatenating,inadequate utilization of abstract features lost in upsampling process,and redundancy of network parameters.These problems are major obstacles in medical image segmentation.This thesis focuses on the key technologies of encoder-decoder architecture in segmentation,the main work is as follows:1)The method of using segmentation networks based on the encoder-decoder architecture for medical image segmentation is studied.Three common improvement directions for encoder-decoder structure are summarized,which provides ideas for the improved design of encoder,decoder and skip connection structures.2)A encoder-decoder segmentation model based on position and channel dual attention mechanism is proposed for medical image segmentation.There is a dependency between the feature objects of medical images,and the existing encoder-decoder architecture does not highlight the correlation of similar features.This thesis embeds a dual attention module between the encoder and the decoder to construct a segmentation model.Through comparative experiments,it proves that the proposed segmentation model can enhance the feature representation and improve the segmentation performance.3)A multi-path upsampling convolutional network model for medical image segmentation is proposed.Three consecutive convolutions and residual connections are used in each layer of the encoder to obtain multi-scale features of different receptive fields.In the skip connection structure,the residual unit(the number of units is the same as the number of downsampling of the layer)is introduced to avoid the semantic gap caused by the direct concatenating of the corresponding features of the contraction path and the expansion path.By changing the single expansion path in the decoder and using multiple up-sampling paths,which are respectively spliced with the features processed by skip connection in the encoder layer,the deep abstract features extracted by the encoder can be fully utilized,and the different scale features of the shallow layer can be merged.Meanwhile,the number of network parameters can be effectively reduced by reasonably removing the convolution operation in the expansion path,and the attention mechanism can be easily introduced after the encoder structure of the model.The comparative experiments show that the proposed network model can effectively solve the shortcomings of the encoder-decoder architecture design,and both the evaluation index and the actual segmentation effect have been improved.4)An auxiliary diagnosis system based on multi-path up-sampling network segmentation model is designed and implemented.The system is used to diagnose retinal hemorrhage caused by diabetic retinopathy.It mainly realizes the function of case management and auxiliary diagnosis,which can help doctors reduce misdiagnosis and missed diagnosis,reduce work burden and improve efficiency.