Click here to

Session: AI Applications in Image Guided Adaptive Radiation Therapy [Return to Session]

Physician Evaluation of Deep Learning-Based Dose Predictions for Head and Neck Radiotherapy

M Gronberg1,2*, S Gay2, B Beadle3, A Olanrewaju2, C Cardenas1,2, R Howell1,2, C Peterson1,2, C Fuller1,2, A Jhingran2, T Netherton1,2, D Rhee1,2, L Court1,2, (1) The University of Texas MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences, Houston, TX, (2) The University of Texas MD Anderson Cancer Center, Houston, TX, (3) Stanford University, Stanford, CA


TH-E-TRACK 4-3 (Thursday, 7/29/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: To predict high-quality, realistic dose distributions for head and neck (HN) VMAT plans that can be used as comparative dose distributions for automated peer review of plan quality.

Methods: We trained a 3D dilated dense-UNet to predict 3D dose distributions for HN patients. The architecture takes a 24-channel input including CT images, target prescriptions, and normal tissue and target contours. 118 VMAT HN plans with 2-3 dose levels were split into train, cross-validation, and test datasets (3:1:1 respectively). The model was trained on random patches using a custom weighted mean squared error loss function, which assigns a customized weight to each contour. The predicted dose distributions were assessed by a HN radiation oncologist using a 5-point scale (1:No edits needed, 2:Minor edits optional, 3:Minor edits needed; 4:Major edits needed, 5:Fails to meet clinical criteria). A score of 1-3 indicates clinical acceptability.

Results: The model predicted dose distributions of comparable plan quality to the clinical plans. The mean dose difference between the predicted and clinical dose distributions of the testing dataset was -0.42±0.28Gy. On average, the predicted target DVH metrics of D1, D95, and D99 normalized to the prescription were within -1.94±1.35%, -0.12±3.55%, and 0.64±6.23% of the clinical plans, respectively. The predicted organ at risk metrics of mean and max doses were within -0.35±1.41Gy and -0.83±1.85Gy of the clinical plans, respectively. The predicted dose distributions on the test set were scored as clinically acceptable for 18/21 patients (score 1:5 patients, score 2:9 patients, score 3:4 patients). The other 3 patients received a score of 4 for poor target coverage and target dose fall-off.

Conclusion: The 3D dilated dense-UNet can predict high-quality, realistic dose distributions for HN VMAT plans. Physician assessment indicates that greater than 85% of predictions are clinically acceptable, supporting the use of dose prediction to QA plan quality.

Funding Support, Disclosures, and Conflict of Interest: Mary Gronberg was supported by funding from the CCTS TL1 Program and the American Legion Auxiliary.



    Dose, Quality Assurance


    TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

    Contact Email