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

Abstract

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