| Image segmentation is one of the important research fields in medical image analysis,and it is also an important part of computer-aided diagnosis,monitoring,intervention and treatment.With the continuous development of medical imaging technology and deep learning,neural networks have achieved great success in the field of medical image segmentation and auxiliary diagnosis.However,current image segmentation algorithms cannot perform accurate semantic segmentation of subtle objects,such as extracting 3D microscopic vessels from cerebrovascular image data is still a huge challenge.There are many problems in the segmentation of cerebral blood vessels:(1)The cerebral blood vessels are complex and tiny,with a diameter of only 1-2voxels,and the image ratio of cerebral blood vessels is about 10:1,resulting in the blood vessels occupying a low proportion of pixels in the image.The segmentation network cannot accurately extract the local features of blood vessels;(2)The geometric structure of blood vessels in the brain is complex,and the local blood vessels cross each other and bend each other,which makes the segmentation of blood vessels susceptible to its influence.In view of the above problems,this paper has carried out the following research:(1)In order to solve the problem that the proportion of blood vessel pixels is low and the local features of blood vessels cannot be accurately extracted,this paper proposes a blood vessel segmentation algorithm based on super-resolution.The algorithm uses super-resolution technology to enlarge the brain image,and increases the blood vessel feature information on the basis of maintaining the blood vessel structure,hence improving the accuracy of blood vessel segmentation.(2)In order to solve the problem of blood vessel deformation after super-resolution,this paper introduces the concept of image gradient,and proposes a super-resolution algorithm based on image gradient.Through gradient transformation,the structural information in the image is extracted and integrated into super-resolution.The network provides structural priors for the generation of super-resolution images.The algorithm can make the generated highresolution images retain effective details and avoid segmentation errors caused by structural distortion.(3)In order to further improve the accuracy of MRA vessel segmentation in brain images,this paper proposes a multi-task learning network,which consists of two branches,a superresolution network and an image segmentation network.While training the super-resolution network,the problems encountered by the segmentation network are fed back to the superresolution network,and the two learn from each other,therefore further improving the segmentation accuracy.Compared with the existing cerebrovascular segmentation methods using the Tube TK dataset,the Dice similarity coefficient of our method is improved by 1.7%. |
