Measuring atrophy by simultaneous segmentation of serial MR images using 4-D graph-cuts

We propose a method for simultaneous segmentation of serially acquired magnetic resonance (MR) images. An existing graph-cuts based algorithm is extended and applied to 4-D images. A probabilistic atlas is generated for each baseline scan by intersubject registration of multiple labeled images. The atlases are used for baseline and aligned follow-up images and are combined with an intensity model to define a weighted graph that connects the timepoints. A minimal cut on this graph yields the segmentation for all timepoints. The resulting segmentations are consistent over time in boundary regions with weak gray scale definition, but reflect atrophy well where the structure boundary is well defined by MR intensity. The hippocampus was segmented in 568 baseline and follow-up images provided by the Alzheimer´s Disease Neuroimaging Initiative (ADNI). The estimated atrophy rates correctly classified AD patients from controls at a rate of 82% (Atrophy rates: AD 3.85%/y.; MCI 2.31%/y.; controls: 0.85%/y.)