Research On Fully Convolution Neural Network For Brain MR Image Segmentation

Magnetic resonance imaging(MRI)of brain provides anatomical images with high contrast for soft tissues and high spatial resolution.,which is a non-invasive imaging technology.Brain development is complex and spaned through childhood and adolescence,it is important to develop quantitative tools for analysis of neurodevelopment at all ages.Brain segmentation in MR images is a central piece of such quantitative analysis tools,studying of normal and abnormal early brain development where accurate tissue segmentation of infant brain images into white matter(White matter,WM),gray matter(Gray mater,GM),and cerebrospinal fluid(Cerebrospinal Fluid,CSF)plays an important role.Recently,deep learning algorithms are gaining popularity in many areas of medical image automatic segmentation.Base on this,this thesis investigates brain MR images segmentation by improved 3D fully convolution neural networks,the data used in the network consist of T1 and T2 weighted brain MR images of 6-month infant.By comparing to the segmentation performance with that of the original method,the results show that the improved method has a better performance in brain MRI segmentation.Firstly,this article briefly analyzes the two major drawbacks of using convolution networks to segment images:at first,fully connected layers require fixed size images,can’t assign a class label to each pixel;secondly,pooling layers reduce the spatial size of the image and lost image information at the same time.To tackle the above issues,we introduce fully convolutional networks and some innovative approachs.Next,we describe the fully convolutional networks architecture which we use in this thesis:1.All layer are convolutional layers;2.Multi-resolution information,concatenate feature maps from shallower layers to the ones resulting from the last convolutional layer;3.Multi-modality and 3D volumes input,multi-modality MR images input sources provide more valuable features information,which can help deal with low contrasts in infant brain MRI.3D data can integrate contextual information from neighbouring slices.Then,to improve the performance of network,adding 1 ×1×1 convolution to deepen the depth of the network layer,and join the dropout to prevent over fit.Finally,the two methods proposed are applied to of the test data set,ITK-SNAP was used to show the 3d image of segmentation results,the experimental results show that improved method has a better performance,results are obtained making comparisons using the Dice Coefficient,MHD and ASD between the Ground Truth and the predicted images.