Semantic Segmentation Of Pulmonary Nodules Based On 3D U-Net

In recent years,semantic segmentation of pulmonary nodules using deep learning has become one of the main research directions in the combination of medicine and computer technology.Doctors can refer to the segmentation results of the computer for auxiliary diagnosis,reducing the difficulty and workload of medical diagnosis.With the continuous development of science and technology,the accuracy of using artificial intelligence method in segmenting lung nodules has been significantly improved.However,there are still many difficulties in practical application.Therefore,this paper aims to solve the problems such as unclear target,rough contour and slow network training speed in pulmonary nodule segmentation.Based on lung CT images in public data sets(such as LUNA16 data set,etc.),this study further studied and improved the full convolutional neural network segmentation algorithm.The main research contents and conclusions of this paper include:(1)Preprocessing of lung CT data sets.Lung CT images with source data format cannot be directly used for training of convolutional neural network.Therefore,all images in the published LUNA16 data set and part of images in the LIDC-IDRI data set are first selected as the initial data set.Then,the features of different Heinz units in different regions of the image are taken into account.Removing the parts that are irrelevant to the diseased area reduces the amount of network computation.Then,the training set,verification set and test set are divided,and the data set required for the training of convolutional neural network is established by the commonly used data enhancement methods such as rotation,scaling and flipping.(2)A 3D U-Net image semantic segmentation algorithm based on Attention structure is proposed.Firstly,this study modified the 3D U-Net network model as the basic framework,and used the dual Attention module,namely the location attention module and the channel attention module,to connect the up-sampling and down-sampling parts of the semantic segmentation network to the corresponding layer,so as to enhance the feature mapping ability.Meanwhile,the improved loss function enables the network to process space and channel information at the same time,improving the accuracy of semantic segmentation.(3)An image semantic segmentation algorithm based on lightweight 3D U-Net is proposed.Due to the complex structure of the network model with the Attention structure added,it does not have good real-time performance.Therefore,this study proposes to first build a sparse network to reduce a certain number of channels so as to reduce the computational amount of the network.Secondly,skip connections between the layers of the network make the network lighter and use depth separable convolution to significantly reduce the cost of network training.Finally,the segmentation accuracy of the model is guaranteed by horizontal connection and pyramid pool module.In this paper,the above two algorithms are tested on LUNA16 and LIDC-IDRI data sets.At the same time,qualitative analysis and quantitative analysis are used to compare with the current mainstream semantic segmentation methods.The results show that the two image semantic segmentation algorithms proposed in this paper achieve good segmentation results in different application scenarios,and provide a certain reference value for future semantic segmentation research.