A Robust Intensity Similarity Measure for Multi-Atlas Segmentation
G Sharp1*, M Peroni2, N Shusharina1, J Shackleford3, P Golland4, G Baroni5, (1) Massachusetts General Hospital, Boston, MA (2) Paul Scherrer Institute, Villigen, Switzerland (3) Drexel University, Philadelphia, PA (4) Massachusetts Institute of Technology, Cambridge, MA (5) Politecnico di Milano, Milan, ItalyTH-C-WAB-3 Thursday 10:30AM - 12:30PM Room: Wabash Ballroom
Atlas-based segmentation is a general approach to automatic segmentation that labels regions of an image based on their alignment to existing structures in an atlas image. The atlas-based approach can be improved by aligning multiple atlases with the target image, and fusing their results. A typical strategy for multi-atlas segmentation is weighted voting that combines structure distance with intensity similarity. This abstract investigates the use of a robust measure for penalizing the similarity of voxel intensities when voting.
Experiments were performed comparing the robust measure, a truncated quadratic penalty, with the more commonly used quadratic penalty. An atlas database of 20 subjects with structures segmented on head and neck CT were evaluated. Training parameters were tuned using leave-one-out cross validation.
Automatic segmentation results were evaluated using the Dice similarity coefficient. The average Dice scores for segmentations produced with a quadratic penalty were 0.78 for brainstem; 0.78 and 0.77 for left and right eye balls; 0.66 and 0.64 for left and right parotids. The average Dice scores for segmentations produced with the truncated quadratic penalty were 0.82 for brainstem; 0.85 and 0.84 for left and right eye balls; 0.74 and 0.73 for left and right parotids.
A robust intensity similarity measure, such as a truncated quadratic penalty, can be an effective approach for improving overall segmentation quality for multi-atlas methods.
Funding Support, Disclosures, and Conflict of Interest: National Institutes of Health