We introduce a novel multi-step approach to improved detection of 3D anatomical point landmarks in tomographic images. Such landmarks serve as important image features for a variety of 3D medical image analysis tasks (e.g. image registration). Existing approaches to landmark detection, however, often suffer from a rather large number of false detections. Our multi-step approach combines an existing robust 3D detection operator with two different novel approaches to the reduction of false detections, and is applied within a semi-automatic procedure allowing for interactive control by the user. Experimental results obtained for a number of different anatomical landmarks of the human head in 3D CT and MR images demonstrate that both automatic ROI size selection and incorporation of a priori knowledge of the intensity structure at a landmark significantly improve the detection performance. The applicability of semi-automatic landmark extraction is thus considerably improved. We also summarize the results of a validation study in which we compare the performance of semi-automatic landmark extraction with that of a (standard) manual procedure for landmark extraction. As an exemplary application, we consider rigid MR/CT registration. The main result of our study is that compared to a purely manual procedure, semi-automatic landmark extraction (a) significantly reduces the elapsed time for landmark extraction, (b) generally yields registration results of comparable quality, and (c) increases the reproducibility of the results.