Room: Osceola Ballroom C
Purpose: To use artificial intelligence to automate clinical target volume (CTV) delineation for oropharyngeal cancer patients and, therefore, reduce inter-physician variability which is the largest source of uncertainty in head-and-neck radiotherapy. This auto-delineation would address a rate-limiting step in fully-automated treatment planning.
Methods: To establish a baseline, we assessed inter-physician CTV delineation variability from a team of sub-specialized head-and-neck radiation oncologists. Then, we developed models to auto-delineate (1) high-risk CTVs using a feedforward network that uses tumor and organs-at-risk information, and (2) low-risk CTVs using a 3D convolutional neural network that uses the CT image and GTV. Each algorithmâ€™s auto-delineations were compared to clinically-used CTVs using overlap/distance metrics. These delineations were then used in our automated treatment planning system to evaluate the impact of CTV delineation on dosimetric distributions.
Results: The measured inter-physician variability resulted in a median Dice Similarity Coefficient (DSC) of 0.75, translating to better agreement than commonly reported in literature. The auto-delineations for high-risk and low-risk CTVs performed well in comparison to their respective clinically-used delineations (median DSC values of 0.81 and 0.82, respectively). Auto-delineated volumes provided improved delineation consistency compared to inter-physician variability (p<0.0001). Dosimetrically, coverage was acceptable (per RTOG 1016 guidelines) for 72% of auto-delineated plans when evaluated on clinically-used targets. The median percent-volume of clinically-used CTVs receiving 98% of the prescribed dose was 98% for auto-delineated CTV plans. We found no significant difference in normal tissue dose metrics between physician and auto-delineated target plans using a paired-Wilcoxon Rank-Sum Test: Brainstem Dmax (p=0.23), cord Dmax (p=0.87), contralateral parotid Dmean (p=0.41), and ipsilateral parotid Dmean (p=0.59).
Conclusion: We demonstrated that artificial intelligence-based algorithms can delineate high- and low-risk CTV with less variability than radiation oncologists. This work addresses a major barrier to fully-automated head-and-neck treatment planning by providing high-quality patient-specific CTVs.