Program Information
Comparison of Atlas Selection and Fusion Strategies for Multi-Atlas Based Segmentation of Head and Neck Structures for Adaptive Radiation Therapy
R Haq*, S Berry , Z Saleh , H Veeraraghavan , Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
SU-F-FS4-6 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: Four Seasons 4
Purpose: Accurate and consistent segmentation of organs-at-risk (OARs) is essential for the safe and effective delivery of radiotherapy. We investigated which combination of three different image features and four atlas fusion strategies would result in the best possible multi-atlas based auto-segmentation (MABAS) of Head and Neck (H&N) OARs on CT images.
Methods: Sixteen H&N contrast-enhanced planning CT images, each including expertly delineated segmentations of nine OARs, were used for atlas construction and testing using a leave-one-out strategy. Image features consisted of CT-intensity, Gabor edges, and Modality Independent Neighborhood Descriptors (MIND). The sum of squared distance (SSD) metric was used to rank atlases by registration accuracy. We employed four different atlas fusion strategies consisting of (a) single Best Atlas (BA), (b) Structure-specific Weighted Voting (SWV), (c) Global image Weighted Voting (GWV), and (d) Majority Voting (MV) using the five best matching atlases. In addition to segmentation, we also implemented an approach to display confidences in the generated segmentations. Dice Similarity Coefficient (DSC) and Hausdorff distance (HD) were used to compare the resulting segmentations from the afore-mentioned fusion strategies against expert manually delineated OARs. Analysis of variance (ANOVA) using DSC was computed to determine performance differences between the afore-said fusion methods. P-values <0.05 were considered significant.
Results: GWV using CT-intensities produced the best segmentation accuracy for all structures (median DSC 0.75, Inter-quartile Range (IQR) 0.63 – 0.83, median HD 10.02 mm, IQR 7.31 – 13.91). All voting-based methods significantly outperformed BA (SWV, GWV, MV p-value < 0.001). ANOVA showed no difference in performance between GWV and SWV (p-value = 0.938) but showed significant differences between GWV and MV methods (GWV vs. MV p-value < 0.001; SSW vs. MV p-value = 0.0031).
Conclusion: GWV atlas fusion using CT-intensity produces accurate and computationally efficient auto-segmentation of H&N OARs using the MABAS scheme.
Funding Support, Disclosures, and Conflict of Interest: This work is supported through a master research agreement between the authors and Varian Medical Systems.
Contact Email: