Brain Magnetic Resonance Image Segmentation Based On Weakly Supervision

Magnetic resonance imaging(Magnetic Resonance Image,MRI)is a modern imaging technique that is widely used to examine internal organs of the human body.Magnetic resonance imaging has been widely used in clinical routine examination and diagnosis,especially the detection and diagnosis of brain diseases.Changes in the volume of brain tissue often predict various diseases,such as brain tumors,white matter encephalopathy,and olive pontine cerebellar atrophy(OPCA).Therefore,accurate and automatic segmentation of brain MRI images is of great significance for clinical applications and scientific research.In recent years,deep learning has been increasingly used in the field of medical image processing and analysis.Among them,the fully convolutional neural network can achieve endto-end image semantic segmentation,so it is widely used in medical image segmentation tasks.However,such methods often require large amounts of labeled data as supervisory information to guide neural network training.However,the labeling of data is generally done by professionals such as doctors and experts,which puts a lot of burden on them.Based on the high similarity between individual brain tissues,this paper establishes a new weakly supervised brain magnetic resonance image segmentation framework.Unlike traditional supervised methods that use voxel-level tags,this article only uses a brain magnetic resonance image with supervoxel-level tags.The weakly supervised segmentation framework proposed in this paper mainly includes three parts,high-confidence seed region generation based on supervoxel matching,brain tissue segmentation network and deep seed region growth.First,based on the high similarity between individual brain tissues,this paper researches a super voxel matching method to achieve matching between the image to be segmented and the reference image,so that the supervoxel of the image to be segmented gets a label.Furthermore,among the supervoxels matched by the image to be segmented,a high-confidence supervoxel is selected to form a seed region.Then,based on the full convolutional neural network framework,this paper studies to establish a brain tissue segmentation network model,using the selected seed area as a supervised label,guiding the network training,and obtaining the category probability of each voxel of the image.After that,this paper used deep seed growth method,based on the current seed area,using the class probability output by the convolution network to update the seed area.Finally,iterate the brain tissue segmentation network training and deep seed region growth in two steps until the seed area covers the entire image and output the segmentation result.In order to evaluate the performance of the weakly supervised brain MRI image segmentation method proposed in this study,this paper conducted experiments on two general datasets of IBSR18 and Brain Web20,and verified them by qualitative comparison and quantitative analysis methods.Experimental results show that the method proposed in this paper can still achieve good segmentation results when using weaker supervision information.