Atlas Ranking and Selection for Multi-Atlas Segmentation
J Yang*, B Beadle, A Garden, P Balter, L Court, The University of Texas MD Anderson Cancer Center, Houston, TXTH-C-WAB-4 Thursday 10:30AM - 12:30PM Room: Wabash Ballroom
Purpose: To improve multi-atlas segmentation by ranking and selecting a subset of optimal atlas candidates resembling the image to be segmented in terms of local anatomy similarity.
Methods: Thirteen planning CT images of head and neck cancer patients were acquired (resolution: 1.0x1.0x2.5mm³) and the esophagus were contoured manually by a head-and-neck oncologist. We randomly chose one of the 13 set of images as the test; the remaining 12 became the atlases. Deformable registration was first applied to transform each of the atlas esophagus contours to the test image. We used Kullback-Leibler divergence to measure the similarity of local intensity histograms between the test image and each atlas, and the measurements were used to rank the atlases. Deformed contours from the most to least similar were added sequentially and the overlap ratio was examined. We identified a subset of optimal atlas candidates by analyzing the variation of the overlap ratio versus the number of atlases. The deformed contours from these optimal atlases were fused to produce the final segmentation.
Results: The rank of the atlases determined by the Kullback-Leibler divergence was consistent with the human visual observation. The overlap ratio analysis identified a subset of 5 optimal atlas candidates automatically. By comparing the segmentation generated from the 5 optimal atlases with the manual contour, it achieved a Dice similarity coefficient of 67.2% and a mean surface distance of 2.3mm, whereas the segmentation generated from all 12 atlases had these values of 31.4% and 4.0mm. Contour fusion relies on contour agreement. If more than half of the atlases produce incorrect deformed contours, it is necessary to identify the optimal atlases to improve the contour fusion process in the multi-atlas segmentation.
Conclusion: Ranking and selecting a subset of optimal atlas candidates using local anatomy similarity improves the multi-atlas segmentation.