Multi-atlas segmentation in head and neck CT scans

Abstract

We investigate automating the task of segmenting structures in head and neck CT scans, to minimize time spent on manual contouring of structures of interest. We focus on the brainstem and left and right parotids. To generate contours for an unlabeled image, we employ an atlas of labeled training images. We register each of these images to the unlabeled target image, transform their structures, and then use a weighted voting method for label fusion. Our registration method starts with multi-resolution translational alignment, then applies a relatively higher resolution affine alignment. We then employ a diffeomorphic demons registration to deform each atlas to the space of the target image. Our weighted voting method considers one structure at a time to determine for each voxel whether or not it belongs to the structure. The weight for a voxel's vote from each atlas depends on the intensity difference of the target and the transformed gray scale atlas image at that voxel, in addition to the distance of that voxel from the boundary of the structure. We evaluate the method on a dataset of sixteen labeled images, generating automatic segmentations for each using the other fifteen images as the atlas. We evaluated the weighted voting method and a majority voting method by comparing the resulting segmentations to the manual segmentations using a volume overlap metric and the distances between contours. Both methods produce accurate segmentations, our method producing contours with boundaries usually only a few millimeters away from the manual contour. This could save physicians considerable time, because they only have to make small modifications to the outline instead of contouring the entire structure.