Research On Medical Image Segmentation Based On Optimized 3D-Unet Network

Medical imaging provides an important basis for the discovery of lesions,precise diagnosis,and personalized treatment.In clinical practice,it is also of important reference for doctors’ auxiliary diagnosis.The analysis of specific and complex lesions before surgery often requires Marking the lesion location and manually segmenting the lesion in the image often takes time and effort.Algorithms based on deep learning can better solve this problem.In order to improve the performance of the model,this article mainly conducts the following two research:(1)The original U-net network is often prone to feature loss due to the stacking of multiple convolutions and downsamplings.In order to allow the semantic network to learn more fine features and have better nonlinear expression capabilities,this article compares the original The 3D-Unet network is improved,and Mutations Res 3D-Unet is proposed.By cascading two 3D-Unets and extending the depth of the overall network,the coding part of the first U-net network is combined with the second U-net network.The decoding part is combined to form a new encoder-decoder structure.At the same time,in order to solve the feature accumulation problem of the deep network,this algorithm replaces the convolution and corrected linear unit(ReLU)in each sub-U-net module with the Recurrent Residual Block module.In order to verify the performance of the algorithm proposed in this paper,we are It was verified on the BraTS 2018 data set.The experiment showed that Mutations Res 3D-Unet has a Dice index of 0.9103,0.8650,and 0.8562 for the complete tumor area,core tumor area,and enhanced tumor area of the brain tumor data,and is relative to the original 3D-Unet.0.8775,0.8170,and 0.8380 have significant performance improvements,indicating the effectiveness of this algorithm.(2)The output feature map of the traditional convolution operator can only extract the local information of the input feature map,and the global information is often easy to be ignored.Therefore,in order to integrate the global information into the output feature map,this paper replaces the upsampling operation of the network It is an upsampling Residual Blocks module based on a global aggregation module.At the same time,in order to allow the network to achieve a relatively good segmentation effect in fewer data sets,this chapter introduces an adversarial training mechanism on the basis of Chapter 3,and inputs the output results of the segmentation network and the gold standard data into the discriminator network and uses The discriminator network discriminates whether the input is the gold standard data or the segmentation result,and uses this to correct the segmentation result of the segmentation network.This article uses 75% of the training set of BraTS 2018 for training and verification on the test set.The complete tumor of the brain tumor The Dice indexes of region,core tumor region,and enhanced tumor region are 0.9164,0.8698,and 0.8647 respectively.Compared with the existing brain tumor segmentation algorithms,this algorithm has strong competitiveness.