Research On Segmentation Algorithm Of Deep Brain Structures Based On Convolution Neural Network

As it is non-intrusive,painless,fast to acquire,magnetic resonance imaging(MRI)of brain has become effective a tool in medical practice for diagnosis,disease followup,treatment evaluation and brain development monitoring.Thalamus,putamen,caudate nucleus,globus pallidus,hippocampus and amygdala are important deep brain structures.Their morphological changes are related to many diseases,such as attention deficit hyperactivity disorder and Alzheimer's disease.Segmentation of deep brain structure in brain MRI is helpful to analyze the changes of various structures,thus helping doctors to diagnose and treat diseases.However,manual segmentation of brain MRI is very time-consuming and the easy to be missegmented due to fatigue,which leads to missed diagnosis and misdiagnosis.Therefore,semi-automatic and automatic segmentation methods have received extensive attention and research.Convolutional neural network(CNN)has shown excellent performance in the field of medical image segmentation,but it has not been solved well on the problems of low accuracy of edge segmentation,low accuracy of small brain structure segmentation and low accuracy of joint segmentation of multiple brain structures.Under the framework of deep learning,this thesis improves the above three problems by integrating medical prior information.Based on this,the main work and results in this thesis are as follows:(1)Aiming at the difficulty of segmentation of brain structure edges,a segmentation algorithm of brain structure based on atlas registration and convolution neural network is proposed.Firstly,the probability map generated by training image is used for rough segmentation,and then the lower confidence part of the probability map is segmented by CNN.When using CNN for fine segmentation,an enhanced feature is proposed,which takes into account the class probability of the neighborhood pixels of the target pixels.The algorithm improves the accuracy of edge segmentation and enhances the label consistency of segmentation results.(2)Aiming at the small size and difficult segmentation of brain structures such as globus pallidus and amygdala,a combined segmentation algorithm of brain structure is proposed.The algorithm uses the anatomical characteristics of the brain structure to segment the brain structures which are more difficult to segment.The algorithm extracts the relative position feature,shape feature and 2.5D gray-scale image block feature of brain structure,and processes them as input of CNN for classification judgment.The algorithm improves the segmentation accuracy of globus pallidus,amygdala and caudate nucleus.(3)Aiming at the low accuracy of joint segmentation of multiple brain structures,an algorithm is proposed to fuse local and global information.Firstly,the algorithm locates the whole region of six brain structures,which narrows the segmentation range,and then jointly segmentes the brain structures.In joint segmentation,a multi-channel and multi-scale convolutional neural network is proposed,which makes the learning process of the network take into account both local and global information,and improves the accuracy of joint segmentation of multiple brain structures.