Reducing False Detections in Extracting 3D Anatomical Point Landmarks

Applying 3D di erential operators to extract point landmarks from medical images generally su ers from false detections. A considerable number of these false detections is caused by neighboring structures that are included in the region-of-interest (ROI) speci ed by the observer. The main contributions of this paper are two di erent approaches to reducing false detections resulting from neighboring structures. First, we present a statistical di erential approach to selecting a suitable ROI size automatically. Second, we propose a di erential approach to incorporating prior knowledge of the intensity structure at a landmark. Also, to cope with anisotropic voxel sizes in estimating partial derivatives, we implemented a computationally e cient scheme based on cubic B-spline image interpolation. Experimental results based on 3D MR and CT images of the human head are presented.