Context-Aware Anatomical Landmark Detection: Application to Deformable Model Initialization in Prostate CT Images

摘要: Anatomical landmark detection plays an important role in medical image analysis, e.g., for landmark-guided registration, and deformable model initialization. Among various existing methods, regression-based method has recently drawn much attention due to its robustness efficiency. In this method, a regression is often trained each predict the location of from any voxel based on local patch appearance, also 3D displacement vector landmark. During application stage, predicted vectors all voxels form field, which then utilized final with voting process. Accordingly, quality field largely determines accuracy detection. However, fields by previous methods are spatially inconsistent 1) within same 2) across different landmarks, thus limiting accuracy. The main reason that landmark, independently, landmarks their estimated independently. To address these issues, we propose two-layer context-aware Specifically, first layer designed separately provide initial second refine them jointly using context features extracted results impose spatial consistency landmarks. Experimental CT prostate dataset show our proposed significantly outperforms traditional classification-based both