4 - Optimal surface segmentation with subvoxel accuracy in spectral domain optical coherence tomography images

Abstract

Numerous methods have been developed for simultaneous segmentation of globally optimal multiple surfaces in images. Optimal surface segmentation is one of the most popular methods employed for such applications and has been widely used in segmentation of various medical image segmentation problems including retinal surface segmentation in optical coherence tomography images of the eye. The nodes in the graph-based optimal surface segmentation method are used to encode uniformly distributed orthogonal voxels of the image volume. Thus, limiting the segmentation method to a single unit voxel accuracy (the distance between two adjoining nodes in the corresponding graph space). Segmentation with subvoxel accuracy is feasible by exploiting partial volume information in the voxels of the image volume, which results in adjoining graph nodes with nonequidistant spacings. This chapter reports a generalized graph-based multiple surface segmentation method with convex priors. The method is capable of segmenting multiple target surfaces with global optimality and subvoxel accuracy. In addition, surface terrain smoothness and is encouraged using the incorporation of convex priors. The proposed method is not limited by equidistant spacing between adjoining graph nodes and allows nonequidistant spacing between the adjoining graph nodes. A displacement field is computed from the original data volume to locate the subvoxel-accurate centers within each voxels. Using the information from partial volume effect in such a manner results in nonequidistant spacing between adjoining graph nodes. The globally optimal solution is obtained by computing a minimum s-t cut on the constructed edge-based graph with incorporation of the required surface constraints. The proposed method was validated on 25 optical coherence tomography image volumes of the retina for subvoxel and super-resolution segmentation accuracy. Our approach can be readily extended to higher-dimensional image segmentation.