Assessment of Contour Fusion Method for Auto-Contouring of Head and Neck CT Images
A Larrue*, T Kadir, M Gooding, MIRADA MEDICAL, Oxford, OxfordshireSU-E-J-93 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
The goal of this study is to assess and compare Majority Voting (MV) and STAPLE contour fusion methods for atlas-based auto-contouring of Head and Neck CT images.
A dataset of ten expertly contoured and independently validated H&N images was used for the evaluation. 31 Structures including the brainstem, parotids, mandible and lymph nodes level had been contoured.
Using a leave-one-out methodology, each test patient was auto-contoured in turn, using all nine remaining atlases with commercial auto-contouring software (RTx, Mirada Medical, Oxford, UK). Contour fusion of the resulting nine contour sets was performed using MV and STAPLE.
Dice Similarity Coefficient was computed for each structure and patient against the ground truth. The result of contour fusion was compared to the performance of a single fixed atlas and to the single atlas giving the best result, which represents the result possible if perfect per-organ atlas selection was achievable. It is included to act as an upper bound on potential single-atlas performance.
STAPLE was slightly more accurate than MV for 23 structures and than the single atlas for all the structures. Compared to best possible atlas, it was more accurate for 12 structures. The mean DSC for each structure was 0.71 for STAPLE, 0.67 for MV, 0.59 for single atlas and 0.72 for best possible atlas. For the brainstem, left and right parotids, lymph nodes levels and mandible, DSC was 0.85, 0.76, 0.75, 0.75, and 0.86 for STAPLE and 0.81,0.72,0.65, and 0.84 for single atlas.
Both contour fusion methods improve over the single fixed atlas for all structures. The performance of STAPLE was slightly better than MV for most structures. STAPLE also produced slightly lower variability across patients. We conclude that contour fusion can improve auto-contouring result, and STAPLE is marginally better than MV in H&N contouring.
Funding Support, Disclosures, and Conflict of Interest: All authors are employees of Mirada Medical. There is no conflict of interest with the research submitted.