Influence of Contouring Style On Accuracy of Atlas-Based Auto-Contouring
G Stevenson1, L Durrant2, C Eccles2, K Chu2, T Kadir1*, M Gooding1, (1) Mirada Medical Ltd, Oxford, UK (2) University of Oxford, Department of Radiation, Oncology & Biology, Oxford, UK.SU-E-J-125 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: This study aims to investigate the impact of atlas-based auto-contouring upon the accuracy of contouring for operators with varying training and familiarity with a particular contouring convention.
Methods: A set of ten head and neck CT datasets were contoured by Contourer 1 and validated by another expert clinician. A single dataset was selected to serve as an atlas to auto-contour the remaining nine cases using commercial deformable image registration software (Mirada RTx, Mirada Medical, UK). For each of these nine cases, the brainstem, spinal cord, all parotid and submandibular glands were independently contoured by Contourers 1, 2 and 3 (a) manually from scratch and (b) by editing the auto-contours. Contourer 2 was familiar with the contouring convention of the atlas whilst Contourer 3 was not.
Dice similarity coefficient (DSC) was used to compare manual contours and the edited contours against the original contours of Contourer 1.
Results: Mean DSC for all structures for Contourers 1, 2 and 3 were 0.89, 0.85 and 0.66 for the scratch contours and 0.87, 0.84 and 0.74 for the edited auto-contours. The greatest effect of using the auto-contours for Contourer 3 DSC occurred for the parotids (0.52 to 0.67) and submandibular glands (0.77 to 0.85) and the least for the brainstem (0.84 to 0.85) and cord (0.55 to 0.55).
Conclusion: Atlas-based auto-contouring using deformable image registration was found to improve the accuracy of Contourer 3 only and only for the parotids and submandibular glands. DSC for Contourers 1 and 2 are largely unaffected by auto-contours. This suggests that familiarity with contouring convention, or lack thereof, is an important factor in assessing auto-contouring efficacy but that auto-contouring may improve accuracy to a standard.
Within a clinical setting, or for clinical trials, atlas-based contouring may have benefit in training and quality assurance in developing consensus.
Funding Support, Disclosures, and Conflict of Interest: G.S,M.G and T.K are employees of Mirada Medical and there was no conflict of interest with the research submitted.