Brain MRI Segmentation Based On Dense U-Net And Its Application

Accurate quantitative segmentation of brain tissue is the premise of brain development evaluation and surgical planning.Many neurologic diseases originated from the brain development period of infants,so it is very important to evaluate the brain development of infants and young children.In the evaluation of brain development and planning of brain surgery,MRI is a very common observation method.Based on MR brain image,the thickness,volume and water content of different brain tissues in infants can be quantitatively analyzed.However,due to the difficulty to keep infant still and the infant brain is immature when taking brain MR image,it brings some inherent problems such as motion artifact,poor image quality,low tissue contrast and fuzzy boundary,especially in the isointensity period,the gray matter and white matter signal intensity of infant brain image is very close,which leads to the low contrast of gray matter and white matter in this stage,which makes the segmentation of infant brain image in isointensity period become a hot problem in the field of medical image segmentation.In this paper,we analyze the reasons for the challenges of isointensity infant brain image segmentation,improve the existing semantic segmentation algorithm,and embed our method into the brain image segmentation platform.The main contents of this paper are as followed:(1)We proposes a new brain MR image segmentation network based on the essentials of dense connectivity and input with multi-modality and multi-stream,which is called Dense U-Net.This network includes three input streams,which are T1 modality stream,T2 modality stream and T1 and T2 forward fusion input stream.In the downsampling stage,the three input streams improve feature reuse and feature complementarity through dense connection structure,so that we can extract multimodal composite information.(2)We analyzes the reasons for the low contrast of the brain image in the isointensity period from the perspective of neuroscience.In the myelination stage of brain development,Schwann cells are added to form the myelin sheath that encapsulates nerve fibers.The main component of Schwann cells is lipids,which makes the composition of white matter and gray matter more similar.Based on the neuroscience analysis and experiment results of segmentation network,we improved the Dense U-Net.In order to improve the extraction of regional structure trend and fine structure features of brain tissue,we replaces the normal three-dimensional convolution kenel by using the combination of three-dimensional dilated convolution kenel with fusion multiple dilation rate in convolution layer,the short and long-range feature information in the large sensing field is considered with less parameters.We trained and validated our model on the dataset of iSeg201 7,an isointensity infant brain image segmentation competition held by MACCAI,then we compared the results with the existing methods,we have achieved excellent performance,especially in white matter segmentation,and reached the highest Dice coefficient.(3)An online brain image segmentation platform is developed,and the proposed brain image semantic segmentation method is applied to the platform.On the one hand,users can segment brain images and visualize the segmentation results on the web expediently.On the other hand,it makes technical pre research for the future development of brain image segmentation method based on interaction.