WE-H-BRC-07: Validation of a Commercial Atlas Based Auto-Segmentation Package For Automated Contour Quality Control




To investigate whether a commercial atlas based auto-segmentation (ABAS) package can be used for automated review and quality control (QC) of physician segmentations, and if so, to determine action levels for future clinical implementation.


42 head and neck cancer (HNC) patient plan CT scans with manually defined expert physician organ at risk (OAR) segmentations were retrospectively identified. Five representative patients were selected as ABAS models and were subsequently applied to each of the remaining 37 studies, resulting in 5 auto-segmentations (AS) per patient. The similarity metric (SM) generated by the ABAS package to describe the anatomical similarity of the CT scan to be segmented against that of the atlas CT was recorded for each AS. A simultaneous truth and performance level (STAPLE) consensus contour was created from the 5 AS on each patient. Each individual AS and the STAPLE were compared to the existing expert manual segmentation using the Hausdorff Distance (HD) and Dice similarity coefficient (DSC).


Software-based SM is not indicative of AS accuracy. STAPLE significantly improved AS quality over the results averaged from individual ABAS for 7 of 8 OARs. DSC values (mean ± 1 standard deviation) for STAPLE compared to the expert segmentations are comparable to the literature for the parotids (0.74 ± 0.10), mandible (0.80 ± 0.07), and submandibular glands (0.68 ± 0.10) but is poor at delineating the brachial plexus (0.32 ± 0.08) and larynx (0.61 ± 0.15).


Automated QC of manually delineated HNC OARs can be achieved with a STAPLE consensus contour constructed from 5 ABAS. Parotid, mandible, and submandibular gland structures can be flagged for manual review and user intervention if the DSC and HD exceed the action levels defined during the commissioning process. Further investigation is necessary to improve the AS results for the brachial plexus and larynx.

Sean Berry, Harini Veeraraghavan, and Margie Hunt hold grants from Varian Medical Systems unrelated to this project