Multi-atlas Based Prostate Segmentation In MR Images

Prostatitis,benign prostatic hyperplasia and prostate cancer are becoming more common in men.Prostate cancer is already the world's second most common men's cancer.Magnetic resonance(MR)can better show the internal structure of the prostate,which is of great clinical significance for the analysis and diagnosis of prostate diseases.The prostate size,shape and location information relative to the surrounding organs and tissues in MR images are clinically important prerequisites for the diagnosis and pathological stage analysis of prostate diseases.At the same time,they also play a key role in guiding the prostate resection and radiotherapy.Therefore,accurate segmentation of the prostate and surrounding organs is essential.However,in the MR images,because of the limitations of imaging technology,the complexity of the prostate MR image itself and individual variation in sizes,shapes and texture information,it is difficult to distinguish with the rest of the surrounding tissue.At present,the segmentation accuracy of computer-based automatic prostate segmentation algorithm is still low and there are still some gaps compared with accurate segmentation.At present,there are various prostate segmentation algorithms in MR images,mainly including pixel-based classification,parametric deformable model,multi-atlas based segmentation and so on.However,the pixel-based segmentation algorithm relies heavily on the performance of the classifier and the extracted feature.And there is a sharp drop in classification accuracy when the data is imbalance;The segmentation algorithm based on the deformable model is sensitive to the position of the initial shape,easily trapped in the local extremum and has a poor robustness and anti-jamming ability;And the multi-atlas based segmentation method makes full use of advantages of high accuracy of manual segmentation and transforms the segmentation problem into the problem of registration.Through the registration technique,it can effectively integrate the prior knowledge of the shape of the manually segmented medical image into the segmentation process and reproduce the segmentation results.Therefore,multi-atlas based prostate segmentation algorithm has become a hot research in recent years.Multi-atlas based segmentation algorithm consists of three major steps:registration,atlas selection and atlas fusion.The atlas selection and atlas fusion are two important points of the study.With a suitable atlas selection and atlas fusion algorithm,the effect of registration errors can be reduced to a certain extent and the segmentation accuracy can be greatly improved.In this paper,we conduct in-depth research into the existing atlas fusion algorithm and propose a novel fusion algorithm based on distance field(DF).The proposed algorithm no longer fuses the label image,but the distance image after distance field transformation for each label image.The DF provides not only the label information but also the distance information to the boundary of the target object.In the proposed method,we assume that image patches from MR and DF images are located on two nonlinear manifolds,and a patch can be linearly represented by its several local neighbors and under the local constraint,the mapping from MR to DF approximates a diffeomorphism.Based on these two assumptions,the fusion weights of the DF patches can be deduced from the weights of the MR patches.We use the sample dictionary to reconstruct the test sample and solve the dictionary coefficient with local anchor embedding,and further use the dictionary coefficient to combine the DF patches to predict a DF image patch for the test sample,and then through weighting average and threshold processing,the label of each point in the MR image is finally obtained.Compared with the popular fusion algorithms,such as Major Voting,Weight Voting,SIMPLE,STAPLE,Nonlocal Patch-based Label Fusion,we find the distance field based fusion algorithm outperforms the popular fusion algorithms;Secondly,in this paper,an ellipsoidal shape priori is also introduced for the case of incorrect segmentation caused by large registration errors.By combining the multi-atlas based segmentation with a priori ellipsoidal shape,a new prostate segmentation in MR images is proposed under the constraint of the ellipsoid prior.Inclusion of this shape prior constraint restricts the regions of interest of the prostate images,which can greatly avoid the interference of the surrounding tissue and organs in the process of atlas selection.In addition,in the process of atlas fusion,the ellipsoidal shape prior constraint can also calibrate and compensate the shape prior obtained by the registration technique,which can effectively avoid the incorrect segmentation caused by the registration error.The proposed method has been evaluated on prostate images of 50 subjects and the experimental results indicate that this algorithm is proven effective and yields mean Dice similarity coefficients of 88.27%.