Multiatlas-Based Segmentation Editing With Interaction-Guided Patch Selection and Label Fusion

摘要: We propose a novel multiatlas-based segmentation method to address the editing scenario, where an incomplete is given along with set of existing reference label images (used as atlases). Unlike previous methods, which depend solely on appearance features, we incorporate interaction-guided constraints find appropriate atlas patches in and derive their weights for fusion. Specifically, user interactions provided erroneous parts are first divided into multiple local combinations. For each combination, well-matched both identified. Then, updated through voxelwise fusion selected derived from distances underlying voxel interactions. Since different combinations used step, our can consider various shape variations during update, even only limited Besides, since does not either image or sophisticated learning steps, it be easily applied general problems. To demonstrate generality method, apply segmentations CT prostate, brainstem, MR hippocampus, respectively. Experimental results show that outperforms methods all three datasets.