Proc. SPIE. 6914, Medical Imaging 2008: Image Processing
KEYWORDS: Image processing algorithms and systems, Liver, 3D image reconstruction, Data modeling, Magnetic resonance imaging, Image segmentation, 3D modeling, Reconstruction algorithms, Statistical modeling, 3D image processing
We present an algorithm for fully automatic reconstruction of 3D position, orientation and shape of the human liver from a sparsely covering set of $n$ 2D MR slice images. Reconstructing the shape of an organ from slice images can be used for scan planning, for surgical planning or other purposes where 3D anatomical knowledge has to be inferred from sparse slices.
The algorithm is based on adapting an active shape model of the liver surface to a given set of slice images. The active shape model is created from a training set of liver segmentations from a group of volunteers. The training set is set up with semi-manual segmentations of T1-weighted volumetric MR images. Searching for the optimal shape model that best fits to the image data is done by maximizing a similarity measure based on local appearance at the surface. Two different algorithms for the active shape model search are proposed and compared: both algorithms seek to maximize the a-posteriori probability of the grey level appearance around the surface while constraining the surface to the space of valid shapes. The first algorithm works by using grey value profile statistics in normal direction. The second algorithm uses average and variance images to calculate the local surface appearance on the fly.
Both algorithms are validated by fitting the active shape model to abdominal 2D slice images and comparing the shapes, which have been reconstructed, to the manual segmentations and to the results of active shape model searches from 3D image data. The results turn out to be promising and competitive to active shape model segmentations from 3D data.