Evaluation of Automated and Manual Contours for Head and Neck Cancers
J S Buatti1*, N Kirby1, R Li1, M de Oliveira1, C Kabat1, N Papanikolaou1, N Paragios2, S Stathakis1, (1) University of Texas Health Science Center, San Antonio, San Antonio, TX, (2) TheraPanacea, Paris, France
Presentations
PO-GePV-M-17 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To evaluate artificial intelligence (AI) generated contours made with ART-Plan software (TheraPanacea, Paris, France) with respect to manually defined contours from previously treated head and neck (H&N) plans.
Methods: 144 patients treated for H&N cancers were chosen for comparing automated and manual contours. 20 organs at risk (OARs) were compared and included the cochleae, eyes, optic nerves, lenses, optic chiasm, parotid glands, brachial plexuses, esophagus, oral cavity, mandible, whole brain, brainstem, spinal cord, and larynx. All 20 OARs were generated on computed tomography images using ART-Plan, a commercially available AI-based auto contouring software. Automated contours were accepted “as is” for comparison. Expert-defined contour sets from previous treatments did not include all 20 OARs. For those contours that were not manually defined, a comparison was not done. Contours were evaluated using Dice similarity coefficient (DSC), mean volumes, the 95th percentile of Harsdorf distances (HD95) and volume difference (VD).
Results: Automated contours were most similar to manual contours for the brain, right eye, left eye, mandible, brainstem, left parotid gland, and right parotid gland with DSC mean and standard deviation of 0.921 ± 0.197, 0.899 ± 0.029, 0.895 ± 0.029, 0.883 ± 0.049, 0.843 ± 0.087, 0.815 ± 0.072, and 0.810 ± 0.072, respectively. The least similar contours were the left brachial plexus, right brachial plexus, left cochlea, and right cochlea with mean DSC below 0.44.
Conclusion: The comparison of automated and manually defined contours demonstrated that the AI-based contouring software may be acceptable for certain OARs but not for all OARs. While DSC is heavily impacted by the total volume of the OARs, additional metrics including Hausdorff distance and volume difference improve the interpretation of software performance.
Funding Support, Disclosures, and Conflict of Interest: Dr. Nikos Paragios discloses that he is an employee and owner of TheraPanacea
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email