Click here to


Are you sure ?

Yes, do it No, cancel

Automatic Detection of Dental Artifact in a Fully-Automated Treatment Planning Workflow

S Hernandez*, C Sjogreen, S Gay, T Netherton, A Olanrewaju, C Nguyen, D Rhee, J Mendez, L Court, C Cardenas, MD Anderson Cancer Center, Houston, TX


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: The Radiation Planning Assistant (RPA), is a web-based tool that will provide automated treatment planning for resource-constrained clinics. As part of this workflow, the user must re-calculate the final dose in their own TPS. We investigated the optimal workflow for auto-planning in the presence of dental artifacts.

Methods: We (1)developed an automatic dental artifact identification tool and (2)assessed its use in an automated workflow. (1)Over 80,000 HN CT slices (549 patients) were manually annotated by three users and majority-voting was applied to define the presence or absence of dental artifact. The patients were sub-divided into train, cross-validation, and test datasets (3:1:1 respectively). Since <4% of slices had artifact, a random subset of CT slices without dental artifact was used to balance the classes (1:1) in the training dataset. The Inception-V3 deep learning model was trained with a binary cross-entropy loss function. (2) Using this model, we investigated various density override methods applied pre- and post-optimization on 15 independent HN CT scans. The effects of methods on D-max, dose to normal structures, and V95 for PTV1 were quantified.

Results: Sensitivity/specificity on a per-slice-basis were 91%/99%. The model identified all patients with artifact and never misassigned artifact. Small dosimetric differences were observed between the various density-override methods (±1%). Applying such methods pre-optimization resulted in an average reduction of 1.5% to D-max. For the pre- and post-optimized plans, 57% and 94%, respectively, of dose comparisons resulted in normal structure dose differences of ±1%. Majority of differences in V95[%] were within ±0.2% for both methods.

Conclusion: Dose differences across multiple density-override methods were small. However, applying dental artifact management before plan optimization resulted in a more appreciable dosimetric impact relative to after. Therefore, we plan to offer RPA users these options to implement their choice of dental artifact management prior to plan optimization.

Funding Support, Disclosures, and Conflict of Interest: Our research group receives funding from the NCI and Varian Medical Systems


Treatment Planning, Image Artifacts, Image Analysis


Not Applicable / None Entered.

Contact Email