Research And Practice Of Brain Image Segmentation Algorithm Based On U-Net++

The precise segmentation of magnetic resonance imaging(MRI)of the infant brain into white matter(WM),gray matter(GM),and cerebrospinal fluid(CSF)is an indispensable basis for medical researchers to study the early development of the brain during Morphological changes and developmental abnormalities that occur in infancy,such as autism,schizophrenia,bipolar disorder and attention deficit disorder,are indispensable for medical researchers to study the morphological changes in the early development of the brain.However,due to the inherent myelination phenomenon in the human brain,infant brains at five to eight months in the isointensity apparatus show extremely low contrast in MRI,which is a great challenge in the task of segmenting MR images of the infant brain at this stage.Early researchers attempted to segment using systems such as SPM,CIVET,and BrainSuite for MRI segmentation of adult brains,but the segmentation results were not satisfactory due to the nature of MRI of brains at the isointense stage.Therefore,how to perform accurate segmentation of MRI of infant brain in isointensity phase is an important direction of current research.Considering the extremely high labor cost required for medical image processing,especially image segmentation tasks,and the rapid development of computer vision technology in the field of auxiliary medical imaging in recent years,numerous model algorithms for medical image segmentation have been proposed,such as U-Net,V-Net,U-Net++and other network structures,segmentation of infant brain MRI using computer vision technology is a highly significant research direction.In summary,this paper presents a study on the computer vision techniques for segmentation of infant brain MRI at isointensity.The specific task objective is to design a medical image segmentation algorithm model that classifies infant brain MRI data pixel by pixel in four categories.Among the existing medical image segmentation algorithm models,UNet++ is undoubtedly one of the networks with excellent performance.Therefore,in this paper,we conduct a study on the MRI segmentation algorithm of infant brain based on the U-Net++ network structure,and design and implement the RU-Net++ network structure based on the UNet++network and the OneSide data enhancement method for infant brain MRI,mainly Two core problems were encountered in the research process:first,the network model structure needs to be optimized to better fit the task of infant brain MRI segmentation and achieve better segmentation results;second,the sample size of the infant brain MRI data set is insufficient,which will affect the training results of the network.To address the problem of poor performance of the U-Net++ network structure in the infant brain MRI segmentation task,this paper tries multiple activation functions,replaces the activation function that is most suitable for the task,and tries to optimize the network performance by adjusting the depth of the network.Through the analysis of experimental results,this paper proposes the RU-Net++ network architecture to solve the problem of low training efficiency and insufficient performance of the network.The experiments demonstrate that RU-Net++reduces the training time by 28.14%and segmentation loss by 12.18%compared with the UNet++network.To address the problem of insufficient number of samples in the infant brain MRI dataset,this paper starts from the traditional data enhancement means to expand the data used for training by data enhancement of the dataset.By experimentally comparing the gains of different data enhancement methods such as additive Gaussian noise and image flipping on the network segmentation results,this paper determines the optimal parameters of various data enhancement methods for this dataset.In addition,for the data enhancement method of image flipping,the analysis of experimental results suggests the effect of MRI stacking characteristics of infant brain on the data enhancement method.Based on this,this paper then proposes a new data enhancement method,OneSide,which exploits the symmetry of brain tissue,and experimentally demonstrates that this data enhancement method can reduce the segmentation error by 10.49%with only 4.27%increase in network training time.