| Medical image segmentation is a key work in many medical applications including 3D reconstruction of medical images,disease diagnosis,surgical planning,medical teaching and so on.Current research on medical image segmentation mainly focuses on automatic segmentation,and Multi-Atlas Segmentation(MAS)has gained wide attention in recent years because of its superior segmentation performance.Label fusion is one of the most important steps in MAS,which has a decisive effect on segmentation results.At present,most of the weighted label fusion algorithms utilize intensity and shape information of atlases and target image to design fusion weights,while the influence of registration is seldom considered.In this paper,we take intensity and registration error information into account simultaneously and propose a new label fusion algorithm.The main work of this paper is as follows:The label fusion algorithm proposed in this paper utilizes registration quality metric in the weight design,which demands the aid of registration quality evaluation algorithm.Therefore,we make a thorough study on registration quality evaluation algorithms.We use the algorithm,Assessing Quality Using Image Registration Circuits(AQUIRC),to estimate registration errors.Using 3D Magnetic Resonance(MR)image data set,we comprehensively analyze the performance of AQUIRC in linear registration error estimation and nonlinear registration error estimation.In linear registration error experiment,we design linear transformation of different types and different deformation degree to simulate registration errors in a multi-circuit image network and analyze the correlation between AQUIRC's estimation and target registration error.The experimental results show that AQUIRC has excellent performance in the estimation of errors introduced by scaling,shear and rotation transformation;nevertheless,it cannot accurately estimate the errors introduced only by translation.In nonlinear registration error experiment,we design locally spherical transformation for simulated error experiments.Besides,real image experiments are also carried out.The results show that AQUIRC can estimate the location and amplitude of nonlinear simulated errors with displacement over 3mm and can correctly reflect the error distribution in real image experiments.We propose a new label fusion algorithm named Registration Error and Intensity Similarity based Label Fusion(REIS-LF),which converts weight calculation into the minimization of difference between segmentation result and true value.We take the correlation between any two atlases into account and use the probability that the two atlases both give wrong label value for fusion weight design.The probability can be jointly given by the registration error and intensity similarity of two corresponding atlas-target image pairs.The registration error is obtained using AQUIRC.Using 3D MR images,we validate our algorithm in brain structure segmentation including hippocampus,thalamus and nuclei of basal ganglia.We use Demons algorithm for nonlinear registration.Comparing the segmentation results of our algorithm with the registration-error-based fusion algorithm named AQUIRC Weighted(AQUIRC-W),we conclude that our algorithm achieves significant improvement over AQUIRC-W in segmentation accuracy and robustness to different image data. |
