Research On CT Image Segmentation Of Spine Based On Deep Learning

The segmentation of spine is a key component in the quantitative analysis of medical images.A good segmentation of vertebra is helpful to the use of computer-assisted medical system and lays a solid foundation for the follow-up spine medical treatment process.Therefore,it is of great significance to study the segmentation method of vertebrae.However,the traditional medical image segmentation technology can no longer adapt current medical clinical development,and has gradually turned to data-driven segmentation methods,of which deep learning technology is one of the methods.it extracts key information from a large number of medical images,and finally obtains a result with higher accuracy than other segmentation methods.It has become the mainstream medical image segmentation method.In this thesis,deep learning segmentation algorithms based on 2D and 3D convolutions respectively are used to conduct research to improve the current medical image segmentation models for current vertebral segmentation tasks.In addition,in view of the difficulty of data labeling in medical image segmentation tasks,a segmentation framework based on semi-supervised learning is built to provide the methods that can obtain better segmentation results with a small amount of labeled data.The main work of this thesis is as follows:(1)Based on the U-net segmentation model,this thesis proposes an improved vertebra segmentation model DAU-net,which effectively improves the segmentation accuracy and reduces the amount of model parameters.Because U-net has insufficient feature extraction capabilities and unable use of multi-scale features sufficient,this thesis proposes an improved segmentation model DAU-net,which improves both the feature extraction module and the branch structure of the decoder.Experiments have proved that from the results of segmentation of vertebrae,the improved segmentation model has an accuracy increase of about 5% compared with U-net;from the perspective of model parameter performance,DAU-net is also in the leading position among all experimental models.Therefore,the effectiveness of this method is verified.(2)This thesis proposes a lightweight 3D convolution operation method,combined with an improved vertebral voxel data preprocessing method,and extends the DAU-net proposed in this thesis per to a 3D model for multi-classification of the spine L1-L5 region,which effectively improves The accuracy of vertebral segmentation also solves the problems of large amount of 3D convolution model parameters and less training efficiency.This method performs convolution operations on the spine voxel data in the plane direction and the space direction respectively.After extracting the features in each direction,the feature maps in the two directions are merged,which effectively reduces the calculation parameters of the conventional 3D convolution operation.In addition,inspired by the current advanced segmentation methods,this article improves the preprocessing method of the vertebra dataset to improve the quality of the dataset.Experiments have proved that the improved 3D DAU-net has effectively improved the segmentation accuracy,and the model parameters are reduced by more than 50%compared with other 3D segmentation models,which solves the problem of difficulty in3 D model training.(3)By constructing a vertebral segmentation framework based on semi-supervised learning to reduce the cost of manual annotation.Although the above supervised learning methods have obtained good segmentation results,they generally rely on high-quality labeled data.However,labeling medical images requires high labor and time costs in reality.In response to this problem,this thesis proposes a vertebral segmentation framework based on semi-supervised learning.This is the first time that semi-supervised learning has been used in vertebral segmentation.It can use a small amount of labeled data to obtain the segmentation results of supervised learning.After comparing the experiments,it is found that the segmentation results based on the semi-supervised learning method as same as the supervised learning method basically,and when the amount of data annotation is 40% of the total,the segmentation result of semi-supervised learning is better than that of supervised learning by about 0.4%,so it is verified the effectiveness of the semi-supervised learning method for vertebral segmentation proposed in this thesis is discussed.