论文部分内容阅读
Purpose: The accurate segmentation of the specific tissues in encephalic image is necessary for the treatment and research of many diseases.Such tasks are usually accomplished by manual contouring,which is time-consuming and unrepeatable.In order to segment the encephalic tissues efficiently and accurately,we study a multi-atlas based segmentation method in this work.Methods: The multi-atlas based segmentation is composed of atlas registration and label fusion.Given the target image to be segmented and a group of atlas images,at the first step,each atlas image is registered to the target image using a deformable registration method.The registration method use cubic B-spline transformation to the model the difference between images and normalized mutual information as similarity metric.At the second step,the transformed segmentations,obtained from the first step,are fused to obtain an estimate segmentation of the target image.The label fusion method use STAPLE(Simultaneous Truth and Performance Level Estimation)to estimate the weight of each atlas through the Expectation Maximum iterating algorithm.And an error correction classifier is trained using corrective learning method to further improve the segmentation results.Results: We applied the proposed method on the 2012 MICCAI Multi-Atlas Labeling Challenge brain image dataset.The result of this method is comparable to other state-of-the-art algorithms.The segmentation get a mean Dice overlap of 0.916590 for 11 target images each with 14 labels.Conclusions: We study a multi-atlas based method to automatically segment encephalic tissues,which features the combination of STAPLE and corrective learning algorithms.It proves to be an effective method.