Improved generation of probabilistic atlases for the expectation maximization classification

Probabilistic atlases present prior knowledge about the spatial distribution of various structures or tissues in a population, used commonly in segmentation. We propose three methods for generating probabilistic atlases: 1) the atlases are constructed in a template space using dense non-rigid transformations and transformed to the space of unseen data, 2) as the method 1 but atlas selection is performed in addition, and 3) the atlases are constructed directly in the space of the unseen data. The methods were evaluated in the segmentation of the hippocampus in 340 images from the Alzheimer's Disease Neuroimaging Initiaitve (ADNI). Dice overlaps (similarity index, SI) were 0.84, 0.85 and 0.87 with reference segmentations and the correlation coefficients for the volumes were 0.84, 0.92 and 0.96 for the three methods tested. Our results show clearly the importance of probabilistic atlases in segmentation.