Image Registration Method Based On ATUM-SEM Image

Reverse engineering the physical structure of nanoscale brain neuron circuit aims to explore the relationship between the neuron circuit and brain function.Recently,due to the rapid development of scanning electron microscope(SEM)with its very high reso-lution.We can get a clearer microscopic structure of nerve tissue.As a method based on sequence slicing,automated tape-collecting ultramicrotome scanning electron mi-croscopy(ATUM-SEM)adopts the roller designed and belt system method for the ultra thin section,and the backscatter detector of conventional field emission scanning elec-tron for imaging.It is a method which is very suitable for large scale statistical and analysis of subcellular structure.The main processes of 3D reconstruction using this method include sample preparation,sample slicing,imaging,image registration,and 3D reconstruction.In this paper,we mainly study the registration method of ATUM-SEM sequence section images,which is an important step in the process of 3D recon-struction.The purpose is to restore the 3D continuity and geometric characteristics of nerve tissue sections and provide a good image data set for subsequent reconstruction analysis.However the biological tissues are soft,and the tissues will be deformed and damaged during the slicing process.In the sample preparation and dyeing process,the sample may be contaminated.Furthermore,there is only certain similarity between adjacent sections which depends on the thickness of sections.The thicker sections will lead to similar similarity of adjacent images.Besides,the X-Y resolution of microscop-ic image can achieves 2 nanometer,but the section thickness is dozens of nanometers,which lead to the anisotropic 3D reconstruction.These factors make the registration of microscopic neural image is a challenging problem.In this paper,we propose a registration method for ATUM-SEM sequence section image.The main processes of our method include:coarse registration,fine registration and image deformation.In coarse registration,we extract the SIFT feature descriptors of the two images and match the feature points.After getting the matching points,we use the least square algorithm to estimate the rigid transformation between the two images and then deform the image to eliminate the rigid deformation between them.In the process of fine registration,we need to eliminate the nonrigid deformation in the image.Because SIFT is a widely used feature descriptor and can express local details in images,so we extract dense SIFT feature points for each image.Then,we use the siftflow algorithm for dense feature point matching.But this matching method does not remove the outliers explicitly.We have designed a new algorithm for outlier elimi-nation.First,we use the support vector regression model with manifold regularization to estimate the transformation function between the corresponding points set,and the transformation function is taken from the regenerated kernel Hilbert space.After we get the transformation function,we want to estimate the probability that each pair is mismatched.Therefore,we use a mixture model and we can find out the posteriori probability of each pair of mismatches by using the EM algorithm and take it as the basis to eliminate the mismatched point pairs.Finally,we use the moving least square method to deform the image.The innovation of this' article mainly includes the following two points.(i)This paper proposes a registration process for ATUM-SEM sequence section image.(ii)Aiming at point set registration problem,we propose a new method of mismatching point pair elimination.The experimental results show that the registration was success-ful.