Image-based deformable models for 3-D automatic segmentation of the brain

The objective of this research is to measure the volume of the human brain from three-dimensional magnetic resonance images using automatic segmentation. We have developed a deformable model segmentation technique which can be used on several medical image modalities. Deformable model segmentation typically involves using a model which approximates the surface of the structure to be segmented, and deforming this model by optimizing a cost function with various energy terms. The preliminary stage of this research is semi-automatic segmentation of the brain. This process involves manual creation of an initial model by placing polygonal approximations of the brain contour on a few MR slices in the image volume and linearly interpolating between them. This results in polygons placed on the rest of the images to complete the initial model creation. We automate our method for measuring brain volume in three-dimensional high resolution magnetic resonance images. The methods published on deformable model brain segmentation have been atlas-based: The brain model is created by registering the brain to a standardized brain atlas. We present a new approach, the "image-based deformable model," in which a model is created directly from the image volumes using histogram classifications and morphological operations. We also present a validation study in which we investigate the accuracy of our automatic segmentation of the human brain on control and patient populations. The validation includes a study of the inter- and intra-rater variability.