3D model-based segmentation of individual brain structures from magnetic resonance imaging data
D. Louis Collins · 1994
This thesis addresses a specific problem of model-based segmentation; namely, the automatic identification and delineation of gross anatomical structures of the human brain based on their appearance in magnetic resonance images (MRI). The approach developed in this thesis depends on a general, iterative, hierarchical registration procedure and a 3-D digital model of human brain anatomy that contains both volumetric intensity-based data and geometric atlas data that co-exist in a brain-based stereotaxic coordinate system. The model contains features derived from an MRI atlas of gross neuroanatomy, that is the result of an intensity average of 305 brains created with an automatic stereotaxic registration procedure developed here. The objective of this thesis is achieved by inverting the traditional segmentation strategy. Instead of matching geometric contours from an idealized atlas directly to the MRI data, segmentation is achieved by identifying the spatial transformation that, under certain constraints, best maps corresponding features between the model and a particular volumetric data set. After automatic recovery of the linear registration transform, the 3-D non-linear transformation is recovered by estimating the local deformation fields, recursively selected by stepping through the entire target volume in a 3D grid pattern, using cross-correlation of invariant intensity features derived from image data. This registration process is performed hierarchically, with each step in decreasing scale refining the fit of the previous step and providing input to the next. When completed, atlas contours defined in the model are mapped through the recovered transformation to segment structures in the original data set and identify them by name. Experiments for registration and segmentation are presented using simple phantoms, a realistic digital brain phantom as well as human MRI data. The algorithm is used to estimate neuro-anatomical variability, to automatically segment cerebral structures and to create probabilistic representations of the same structures. Validation with manual methods shows that the procedure performs well, is objective and its implementation robust.