Research On 3D Biomedical Image Semantic Segmentation Algorithms In Deep Learning

With the further maturity of deep learning algorithms,the increase of computing power,and the continuous accumulation of data,deep learning has developed rapidly in the field of semantic segmentation of biomedical images.3D convolutional neural networks have become the mainstream choice in 3D biomedical image semantic segmentation because of their 3D context information mining capabilities.However,the 3D convolution kernel will greatly increase the number of trainable parameters.Considering that in the field of biomedical images,the available training set is often very limited,which will exacerbate the risk of overfitting the 3D network.Therefore,a trade-off must be made between the network size and its feature expression ability.Due to the limitation of the parameter amount,3D U-Net with simple and efficient network structure is widely used.However,the simple structure also causes the 3D UNet network to be too shallow,the feature expression ability is limited,and the network cannot obtain a large receptive field.At the same time,more down-sampling layers make 3D U-Net lose some details and global position information.Based on the above analysis,this paper conducts research based on 3D convolution kernel parameter reduction and 3D U-Net network optimization.The main research contents and innovations of this paper are summarized as follows:(1)Aiming at the problem of too many 3D convolution kernel parameters,this paper proposes a 3D densely separated convolution(3D-DSC)module to replace the original 3D convolution kernel.The 3D-DSC module is composed of 1D convolution kernels in three directions.At the same time,by further introducing the dense connection between the additional non-linear activation layer and the 1D convolution kernel,the non-linearity of the network can be significantly improved while maintaining a compact architecture.Linear feature expression ability.The simple 3D classification network constructed in this paper based on 3D-DSC has reduced the amount of parameters by 68.3% compared with the normal 3D network of the same depth.It has achieved a classification accuracy of 76.70% on the MRI diagnosis dataset of ADHD brain,compared with the same depth.The normal 3D network improved by 10.8%,and the classification accuracy of 88.42% was obtained on the CT image diagnosis data of liver cancer.Compared with the normal 3D network of the same depth,it improved by 15.4%.(2)Aiming at the problem that the 3D U-Net network is too shallow and the feature expression ability is limited,this paper uses the 3D-DSC module to deepen the 3D U-Net,and proposes a deeper 3D U-Net-38.U-Net has been deepened from 18 layers to 38 layers,which improves the network's feature expression ability and the network's receptive field.In this paper,we obtained a segmentation accuracy Dice Score of 0.8932 on the brain tumor MRI dataset,which is 2.5% higher than 3D U-Net,and a segmentation accuracy Dice Score of 0.8502 on the kidney tumor CT image dataset,which is improved compared to 3D U-Net It's 2.4%.(3)Aiming at the problem of loss of some details and global location information,this paper proposes branch cross-layer dense hollow space pyramid pooling(BA-DenseASPP).On the basis of traditional DenseASPP,BA-DenseASPP introduces a full-resolution branch from the network input and combines it with DenseASPP through cross-layer connections.The fullresolution branch provides lossless details and global position information,so BA-DenseASPP can make up for lost 3D U-Net details and global position information,and DenseASPP also provides larger receptive fields and multi-scale information.In this paper,the 3D U-Net-38 with BA-DenseASPP achieved a segmentation accuracy of 0.9027 on the brain tumor MRI dataset,which was 1.1% higher than 3D U-Net-38 without BA-DenseASPP.It was obtained on the kidney tumor CT dataset The segmentation accuracy Dice Score is 0.8591,which is 1% higher than 3D U-Net-38 without BA-DenseASPP.