Brain Image Parcellation Using Fully Convolutional Neural Network And Multi-Atlas

Since the development of medical equipment and the reduction of cost,medical image analysis plays an increasingly important role in the diagnosis of diseases,diagnosis and treatment plan planning,disease monitoring and so on.In the medical image analysis,brain image parcellation has been intensively studied.Accurate labeling of anatomical region of brain is required in quantitative analysis of magnetic resonance(MR)brain image.In clinical trials,the regions of interest are usually labeled in brain images manually,which is labor-intensive and low reproducibility.In recent years,many studies have proposed automatic parcellation methods for brain image parcellation,among them the multi-atlas based brain image parcellation(MAP)is the most successful method and has been widely adopted.The main idea of MAP is to register multiple brain atlases,which contain manually labeled brain regions,to a new brain image to be parcellated.After registration,manually labeled brain regions can be propagated to the new brain image and fused into the final parcellation result.Obviously,image registration plays a key role in MAP.However,image registration is sensitive to image quality and inter-subject variation of brain structures,and is very time consuming.To tackle the above mentioned problems,this thesis proposes a MR brain image parcellation using full convolutional neural network(FCN)and multi-atlas instead of image registration.We have proposed two FCN based MR brain image parcellation methods:(1)A Generative Adversarial Networks(GAN)with manually selected brain atlases is proposed in this paper to achieve automatic and robust brain image parcellation.In the GAN of this paper,a 3D fully convolutional neural network with multi-level feature skip connection is used as the generative model,and a 3D convolutional neural network is used as the discriminative model.The experimental results demonstrated that the GAN of this paper can achieve more accurate brain image parcellation than MAP.Furthermore,our method is more robust and efficient than MAP.(2)In order to fully utilize the information in multi-atlas.This paper integrate the SE module proposed in SENet(Squeeze-and-Excitation Networks)into our previous GAN framework,by which the multi-atlas features can be automatically and adaptively selected by the networks and the possible errors caused by manual selection of multi-atlas in the previous work can be eliminated.In the experiment,the new GAN with SE module outperforms our previous GAN in parcellation accuracy.