The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-SIPI-277

“Atlas Guided Deformable Models for Automatic Anatomic Labeling of Magnetic Resonance Brain Images”

by Stephanie Roxane Sandor

December 1994

In this work we describe the image processing and analysis techniques used for a special purpose vision system that analyzes images of the human brain. We develop a computerized method to automatically recognize and label anatomic regions in magnetic resonance brain images. Critical to our method is the assumption that variations in normal brain anatomy from subject to subject can be accounted for by global scaling and local shape differences. We then take advantage of significant a priori knowledge about normal brain anatomy and label a subject's MR image by matching this image to an anatomic model of a normal brain.

The approach we take is to model a pre-labeled brain atlas as a physical object and give it elastic properties, allowing it to warp itself onto regions in a subject's brain image. We use deformable models which are energy-minimizing elastic surfaces that can accurately locate image features. We parameterize deformable models with 3-D bicubic, B-spline surfaces and label model points corresponding to lobe regions and major cortical fissures. We design an energy function whose minima occur such that cortical fissure points on the model are attracted to fissure points on the image, and the remaining model points are attracted to the brain surface. A conjugate gradient numerical method minimizes the energy function, allowing the model to automatically converge to the brain surface.

Because deformable models perform their task by fitting themselves to local image features, these models do not perform well in the presence of complex or neighboring image features. Therefore, preprocessing is required in order to derive any meaningful results from deformable model matching. Our procedure matches a model to an image which has been filtered with a 3-D Marr-Hildreth edge detector and a series of morphological transformations. We have developed a morphological algorithm which automatically extracts a brain as a binary object from raw MR images, finds the cortical surface, and identifies portions of the surface which correspond to cortical fissures. Also, the morphological transformations smooth over convoluted openings in the brain surface, decreasing the complexity of the surface to which the deformable model must converge. Once the atlas is deformed to the smoothed brain surface, the original surface is labeled by propagating the anatomical labels from the atlas.

To download the report in PDF format click here: USC-SIPI-277.pdf (18.9Mb)