Brain Contour Refinement

Many contour refinement algorithms are based on the active contour model algorithm, or snakes [20,23], described in Chapter 8. Given an initial estimate of an object boundary, active contours approach the actual boundary by solving an energy minimization problem. Different enhancements are required for different application domains. One enhancement to the classic snakes algorithm for contouring the hippocampus in MR images is described in [27]. For brain images, Vaillant and Dzavatzikos propose the use of a new active contour algorithm for modeling the brain cortex as thin convoluted ribbons embedded in three dimensions, which is specifically designed for mapping the human cortex [33]. Their results are promising, particularly for tracking into cortical convolutions in high-resolution MR images. However, all these methods require a good initial contour.

Aboutanos and Dawant [1] describe a geometric deformable model used to refine an initial brain mask. Their deformable model uses the pixel intensity along lines that are placed approximately perpendicular to the initial contour. A five-term cost matrix is associated with transforming the image to hug the contours; in addition, a sixth term is required to repel the optimum curve from image locations such as eye and skin locations in T1-weighted images. The authors have found values for these parameters that perform well on sagittally displayed brain contours of 3D T1-weighted MP-RAGE volumes on many volunteers, although the method requires a very good initial contour and excess fat can affect results. Two iterations are applied, and the blurred image is used to reduce the effect of noise. This method looks very promising for T1 images, but no results are presented for PD- and T2-weighted images.

0 0

Post a comment