Click here to


Are you sure ?

Yes, do it No, cancel

Interactive Refinement of Head and Neck Structures Using Multi-Atlas Based Probabilistic Segmentation Maps

R Haq*, S Berry , A Iyer , A Apte , M Hunt , J Deasy , H Veeraraghavan , Memorial Sloan-Kettering Cancer Center, New York, NY


(Tuesday, 7/31/2018) 4:30 PM - 5:30 PM

Room: Exhibit Hall | Forum 2

Purpose: To generate probabilistic OAR segmentations in head and neck CT images by using multi-atlas based auto-segmentation (MABAS) to enable user-interactive refinement of auto-generated contours.

Methods: We developed a novel technique to generate interactive OAR segmentations from probabilistic segmentations computed from MABAS. Our approach automatically selects the appropriate number of atlases required for multi-atlas fusion to each new target image and weights each atlas contribution using image similarity-based measures. Selected atlases were combined with voxel-wise weights to generate probabilistic segmentations with label confidence ranging from 0 to 100% that indicate the voxel-wise confidence in the segmentation. Selecting the appropriate confidence level interactively generated OAR segmentations. Patient-specific structure confidence maps were generated for the left and right parotids, brainstem, mandible and spinal cord. Contours for fifteen patients were refined by changing the confidence level, and compared against expert clinical segmentations. We determined the efficacy of using confidence maps for refinement by comparing quality metrics achieved using confidence segmentation versus MABAS using a default label confidence threshold of 33%. Wilcox rank-sum test was performed to assess the difference between interactive and auto-generated contours using segmentation metrics.

Results: Improvement in contour accuracy was observed the most in brainstem in 66% of the cases, with statistically significant improvement observed for cord (p-value<0.001 for 0.85 vs 0.83 DSC, 2.76 mm vs 3.30 mm 95th percentile Hausdorff Distance) where 60% of the cases required modification.

Conclusion: Probabilistic segmentations provide a visual assessment of the underlying variability within a generated segmentation, and can be used to interactively select between different segmentations. Our preliminary results suggest the feasibility of interactively refining auto-generated contours without requiring the manual editing of the OARs.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by Varian Medical Systems


Segmentation, Image Analysis, Contour Extraction


IM/TH- image segmentation: CT

Contact Email