Research On Atlas-Forest-Based Human-Brain MR Image Labeling

Brain tissue segmentation in MR images has been widely used in medical diagnosis and research.Traditional manual labeling and semi-automatic segmentation methods require manual intervention,which are time-consuming,labor-intensive and subjective.Atlas-based methods combine the processing capability of natural image segmentation methods and the prior knowledge provided by the experts’ manual labels.Furthermore,the multi-atlas-based methods integrate the results of several single atlas-based method to improve the accuracy of segmentation.Base on those reasons,it is widely acknowledged that the atlas-based method is one of the most effective automatic segmentation approaches.In general,the traditional multi-atlas based methods include two main steps: registration and label fusion.Registration is to align the atlas to the target image space so that the prior label information of atlas can propagate to the target image directly.Label fusion is to integrate the segmentation propagated by atlas to the result.In the atlas-based method,registration is a crucial factor and inaccurate registration will recede the credibility of label propagation.However,an atlas-forest-based automatic segmentation method can calculate result without the influence of registration.For the reason that the atlas-forest-based method encodes each atlas into an atlas forest through the random forest,which integrate the characteristics of target,such as intensity,texture and spatial location of interested tissues,and predicts the target through the trained learning models.In this way,the atlas-forest-based method can save the computation time for registration.An automatic segmentation method through extensible learning for atlas selection is proposed to solve the problem that dealing with the dataset that has a large number of atlases.Apart for the traditional atlas selection,the extensible learning method established a selective prior list for atlas in the dataset and updated the list during labeling.This strategy improves the efficacy of atlas selection with the increased number of labeled target image,which makes it possible to select similar atlas for target without traversing the dataset,so that improve the segmentation accuracy dynamically.An automatic segmentation method through spatial indexes based atlas forest is proposed to solve the problem that the traditional multi-atlas based method only considers the information of the atlases itself but does not take into account the association between the atlases.This method puts all the atlas patches into a samples pool and establish the relationship between the atlases through the spatial indexes.In addition,this method samples the patches in the samples pool by the spatial indexes,which reduces the redundancy of the samples and improves the speed of the algorithm.Furthermore,the proposed method uses strategy of random forest to train the samples pool to generate a hybrid atlas forest for integrating the image information of the atlases and the relationship among the atlases in the dataset.Experiments show that the hybrid atlas forest improved the accuracy of segmentation without the influence of registration.An automatic segmentation method through confidence probability based hybrid atlas forests is proposed to solve the problem that the difficulty in evaluating confidence propagated by each voxels.This method trains the atlas dataset as a whole using the hybrid atlas forest and treats each voxel differently through confidence-weighted probability matrix.Experiments indicates that the proposed method can integrate the prior information and correlation among the atlases to obtain an accurate segmentation.An automatic segmentation method using the hashing retrieval based hybrid atlas forest is proposed to solve the problem that the information of target image is not fully utilized.This method considers each pixel in the atlases as a sample and encodes the samples with locally sensitive hashing for fast data retrieval.Each target tissue is encoded into a hash table and the samples are retrieved depending on the similarity to target.Then,a hybrid atlas forest is trained using these selected samples,which is used to predict the labels of the target.The method construct a target-oriented learning model by integrating information among the atlases to improve the accuracy of segmentation.These automatic segmentation methods proposed above can solve the problem of fast and dynamic atlas selection through an extensible learning model and the problem of how to remove the redundant of samples,how to evaluate the brief of label propagation and how to combine the target-oriented thought to segmentation through a hybrid atlas forest.The experiment results shows that these proposed methods can improve the accuracy and efficiency of current brain segmentation methods.