Segmentation of MR Brain Scans Using Atlas-Guided Deformable Contours

Applications of Computer Vision in Medical Image Processing

Volume

Papers from the 1994 AAAI Spring Symposium

Contents

Our approach to segmentation differs from the more common strategies that seek to classify individual pixels based on the measurements at that pixel and in a small neighborhood surrounding that pixel [2]. Instead, we use a boundary finding paradigm, common in computer vision, in which energy minimizing deformable contours (snakes) corral the region [3]. In many snakes applications, the underlying assumption of segmentation is that the "objects" are delineated, along a closed-contour locus, by edges or some other image feature. Of course, the snake may be confused by gaps in the boundary due to low-contrast regions, or by noisy areas where an edge is hard to locate. Regularizing terms can compensate for noise and irregularities by smoothing the contour so that it jumps the gap in troublesome regions. Such fixes are, in general, unreliable, however.