@MASTERSTHESIS\{IMM2014-06748,
    author       = "C. B. Anker",
    title        = "Segmentation of Subcortical structures in T1 weighted {MRI} as a component of a Brain Atrophy Computation Pipeline",
    year         = "2014",
    school       = "Technical University of Denmark, Department of Applied Mathematics and Computer Science",
    address      = "Richard Petersens Plads, Building 324, {DK-}2800 Kgs. Lyngby, Denmark, compute@compute.dtu.dk",
    type         = "",
    note         = "Supervised by: Prof. Mads Nielsen, {KU,} Prof. Rasmus Larsen, rlar@dtu.dk, {DTU} Compute, Prof. Knut Conradsen, knco@dtu.dk, {DTU} Compute, \& Postdoc Mark Lyksborg, {DTU} Compute",
    url          = "http://www.compute.dtu.dk/English.aspx",
    abstract     = "Among the top performing automated hippocampal segmentation methods from structural Magnetic Resonance Imaging (MRI), are multi-atlas segmentation methods, which rely on manual annotations.
In this thesis two fundamentally different multi-atlas segmentation methods are implemented, {N-L} Patch and BrainFuseLab. In {N-L} Patch, each voxel is segmented using information from atlases which have been coarsely aligned using affine registrations. BrainFuseLab aligns atlases using non-rigid registrations, and is thus comparatively slower. To make a fair comparison, both methods will use the same atlases from a new Harmonized Hippocampal Protocol (HHP).
Method parameters are optimized in a leave-one-out cross-validation using two different atlas sets. Based on volume overlap with the manual annotations, {N-L} Patch is chosen to segment a standardized {ADNI} dataset containing 1.{5T} MRIs from 504 diagnosed subjects (169 cognitively normal (CN), 234 mild cognitive impairment (MCI), 101 alzheimer's disease (AD)) at baseline, month 12 and month 24. Hippocampal atrophy calculated as percentage volume change from baseline to follow-up is estimated. Based on a statistical analysis, the diagnostic group separation capabilities of {N-L} Patch are compared to two state-of-the-art methods, cross-sectional FreeSurfer and longitudinal FreeSurfer.
Including the {HHP} annotations in {N-L} Patch yielded signicantly better group separation than cross-sectional FreeSurfer in separating {AD} from {CN} and {AD} from {MCI}. This illustrates the longitudinal robustness of segmentations when annotations from the new hippocampal standard are included in automated segmentation methods. Also longitudinal FreeSurfer exploiting baseline and follow-up simultaneously showed no diagnostic improvement over {N-L} Patch."
}