Click here to

Session: Machine Intelligence for Treatment Planning and Segmentation [Return to Session]

Deep Learning-Based Dose Prediction for Automated, Individualized Quality Assurance of Head and Neck Radiotherapy Plans

M Gronberg1,2*, S Gay1,2, B Beadle3, A Garden1, H Skinner4, T Netherton1,2, W Cao1, I Vazquez1, A Olanrewaju1, C Chung1, C Cardenas5, C Fuller1,2, C Peterson1,2, A Jhingran1, R Howell1,2, T Lim1,2, M Yang1,2, R Mumme1, L Court1,2, (1) UT MD Anderson Cancer Center, Houston, TX, (2) MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences, Houston, TX, (3) Stanford University, Stanford, CA, (4) University of Pittsburgh, Pittsburgh, PA, (5) University of Alabama at Birmingham, Birmingham, AL

Presentations

WE-G-BRC-3 (Wednesday, 7/13/2022) 2:45 PM - 3:45 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Planners of all experience levels may find themselves unsure if an optimal plan has been achieved. The purpose of this study was to use deep learning-based dose prediction to assess the quality of head and neck (HN) plans and identify sub-optimal plans.

Methods: A dataset of 245 VMAT HN plans was created using RapidPlan knowledge-based planning (KBP). A subset of 112 high-quality plans was selected under the supervision of a HN radiation oncologist. We trained a 3D dilated dense-UNet to predict 3D dose distributions using three-fold cross-validation on 90 plans, using CT images, target prescriptions, and normal tissue and target contours as model inputs. We tested the model’s performance on a hold-out set of 22 plans. We then tested the application of the dose prediction model for automated review of plan quality. Dose distributions were predicted on 14 clinical plans. A comparison of the predicted versus clinical OAR dose metrics was used to flag OARs with sub-optimal normal tissue sparing using a threshold of 2Gy dose difference or 3% dose-volume difference. OAR flags were compared to manual flags of 3 HN radiation oncologists.

Results: The predicted dose distributions were of comparable quality to the KBP plans. On average, the predicted target DVH metrics of D1, D95, and D99 were within -2.53±1.34%, -0.42±1.27%, and -0.12±1.97%, respectively. The predicted OAR mean and maximum doses were within -0.33±1.40Gy and -0.96±2.08Gy, respectively. For the plan quality assurance study, radiation oncologists flagged 37/167 OARs for plan improvement. There was high inter-physician variability; 89% of physician-flagged OARs were flagged by only 1/3 physicians. Our comparative dose prediction model flagged 58/167 OARs, including 19/37 physician-flagged OARs.

Conclusion: Deep learning can be used to predict high-quality dose distributions, which can be used as comparative dose distributions for automated, individualized assessment of HN plan quality.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Cancer Prevention and Research Institute of Texas, the American Legion Auxiliary, and the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Numbers TL1TR003169 and UL1TR003167.

Keywords

Quality Assurance, Treatment Planning

Taxonomy

TH- External Beam- Photons: Development (new technology and techniques)

Contact Email

Share: