Performance of 3D landmark detection methods for point-based warping in autoradiographic brain imaging

Warping can be used to reduce interindividual structural variations of three-dimensional anatomical image datasets of brains. For this, a standard brain is generated and the individual datasets are matched to this reference system. Model-based warping uses structural information such as landmarks to construct the spatial correspondence between the datasets. In this paper we compare three different fully automatic landmark detection algorithms. The first two approaches use a threshold-based definition of landmarks. In the first case (Monte-Carlo based) searching points move in subvolumes simultaneously. In the second approach the searching points move along "rays". The third method uses spatial derivations of voxels to determine the position of landmarks at prominent structural features. The subsequent warping is based on a distance-weighted method with an exponential weighting function. All methods tested are able to reduce structural variations. Best results are obtained by the derivation approach.