Designing and selecting features for MR image segmentation

Deformable models have obtained considerable success in medical image segmentation, due to its ability of capturing the shape variation of the target structure. Boundary feature is used to guide contour deformation, which plays an decisive part in deformable model based segmentation. However, it is still a challenging task to obtain a distinctive image feature to describe the boundaries, since boundaries are not necessarily in accordance with edges or ridges. Another challenge is to infer the shape for the given image appearance. In this paper, the anatomical structures from MR images are aimed to be segmented. First, a new normal vector feature profile (NVFP) is employed to describe the local image appearance of a contour point formed by a series of modified SIFT local descriptors along the normal direction of that point. Second, the shape of the target structure is inferred by matching two image appearances of the test image and learned image appearance. A new match function is designed to incorporate the new NVFP to deformable models. During the optimization procedure of the segmentation algorithm, the nearest neighbor approach is used to compute the displacement of each contour point to guide the global shape deformation. Experimental results on prostate and bladder MR images show that the proposed method has a better performance than the previous method.