SU-E-J-101: Weighted Voting Method for Multi-Atlas Segmentation in CT Scans

Abstract

Purpose: We investigate automating the task of segmenting structures in head and neck CT scans, to minimize time spent manually contouring. We focus on the brainstem and left and right parotids.

Methods: To generate contours for an unlabeled image, we assume an atlas of labeled 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 targetimage. Our weighted voting method acts one structure at a time to determine for each voxel whether or not it exists in a structure. The weight for a voxel's vote from each atlas depends on the intensity difference of the target and the transformed atlas at that voxel, in addition to the distance of that voxel from the boundary of the structure.

Results: We applied our method to sixteen labeled images, generating automatic segmentations foreach using the other fifteen images as the atlas. We compared the resulting Dice and Hausdorff metrics with a majority voting method using the same registrations and saw remarkable improvement. Mean Dice scores were around .7, with maximum Hausdorff of about 15mm, and mean Hausdorffs around 2 or 3mm.

Conclusions: Our method produces contours with boundaries usually only a few millimeters away from the manual contour, which could save physicians considerable time, because they only have to make small modifications to each slice instead of contouring from scratch.

Keywords: Computed tomography; Image transforms; Medical image segmentation; Medical imaging.