@MASTERSTHESIS\{IMM2007-05307,
    author       = "B. S. Dissing",
    title        = "Segmentation Of Regions In The Medial Temporal Lobe",
    year         = "2007",
    school       = "Informatics and Mathematical Modelling, Technical University of Denmark, {DTU}",
    address      = "Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby",
    type         = "",
    note         = "Supervised by Assoc. Prof. Rasmus Larsen, {IMM,} {DTU}.",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/5307-full.html",
    abstract     = "This thesis describes a method used to construct a segmentation algorithm for the Medial Temporal Lobe (MTL) in the human brain. The regions of interest in the {MTL} consists of four closely connected cortical areas, hippocampus and amygdala. A set consisting of 13 {MRI} scans of different individuals with matching manual expert annotations of the {MTL} were delivered, and the intensity images was standardized across all 13 images to get corresponding intensity values for corresponding tissue types.

The regions of the {MTL} are located fairly deep in the brain where the contrast is quite low, and region-boundaries can be diffecult to find. Many classical segmentation approaches will fail to segment these regions of interest, and the key to a more successful segmentation is incorporation of prior knowledge. The manual annotations are represented as Level Set Functions, and a coupled statistical shape model is trained to capture the spatial variation across the dataset. This model is employed in a region-based energy formulation which maximizes the mutual information between the region labels and the image voxel intensity values of the image. A rather signicant improvement is seen when the coupled model is compared to individual segmentations of each region."
}