Atlas image labeling of subcortical structures and vascular territories in brain CT images

We propose a multi-atlas labeling method for subcortical structures and cerebral vascular territories in brain CT images. Each atlas image is registered to the query image by a non-rigid registration and the deformation is then applied to the labeling of the atlas image to obtain the labeling of the query image. Four label fusion strategies (single atlas, most similar atlas, major voting, and STAPLE) were compared. Image similarity values in non-rigid registration were calculated and used to select and rank atlases. Major voting fusion strategy gave the best accuracy, with DSC (Dice similarity coefficient) around 0.85 ± 0.03 for caudate, putamen, and thalamus. The experimental results also show that fusing more atlases does not necessarily yield higher accuracy and we should be able to improve accuracy and decrease computation cost at the same time by selecting a preferred set with the minimum number of atlases.