Brain Image Parcellation Using Fully Convolutional Network With Adaptively Selected Features From Brain Atlases

With the continuous optimization of computer and medical imaging related methods,a large number of researchers have begun to use medical image analysis technology for theoretical research and clinical diagnosis.In the field of medical image processing,in order to make correct treatment decisions for specific internal tissues and organs,medical researchers often need to fully understand the specific situation,and then they must use medical image parcellation.The parcellation and labeling of brain image regions is an important data processing step in neuroscience,and is often used for quantitative analysis of magnetic resonance brain images(MR Brain Image).Since multi-atlas based parcellation(MAP)uses prior information from brain atlases(i.e.,manually labeled brain regions),it can provide accurate brain parcellation and has been widely adopted.Recently,some researchers have used the deep learning based brain image parcellation(DLP)method based on the fully convolutional network(FCN).Compared with MAP,DLP has high computational efficiency,making it more applicable in practice.However,existing DLP methods either neglect or partially utilize brain atlases,making it difficult to get comparable parcellation accuracy as MAP.In this thesis,we propose a new DLP method so as to solve the above problems,which can effectively use the brain atlas.(1)The network is based on FCN and nonlocal channel attention mechanism module NL Module(Non-local Module).The input of the network is the target brain atlas to be parcellated and all brain atlas for auxiliary parcellation.The main parcellation network integrates the brain atlas features selected by the NL module at different network levels as a guide and generates the final parcellation result.(2)Through experiments,it is found that the effect of using all brain atlas information is not as good as using partial atlas.In this thesis,the brain atlas selection was performed in the last step of the NL module,and several brain atlases with higher weights were selected to be integrated into the main network,and obtained better parcellation effect.In experiments on using two public magnetic resonance brain image datasets(LONI LPBA40 and NIREP-NA0),our method is superior to MAP and the latest DLP method due to the effective use of brain atlas.