2294 - Machine Learning Dose Prediction for Automated Proton Therapy Planning in Esophageal Cancer

02:30pm - 04:00pm PT

Hall F

Screen: 29

POSTER

Presenter(s)

Markus Wells, MD, MS - Medstar Georgetown University Hospital, Washington, DC

M. Zarenia, M. T. Wells, M. Shang, P. Jermain, T. Kearney, A. Balawi, K. R. Unger, S. Rudra, and D. Pang; Department of Radiation Medicine, MedStar Georgetown University Hospital, Washington, DC

Purpose/Objective(s): To investigate the applicability and accuracy of a deep learning (DL) model for prediction of radiation dose distribution for esophageal cancer patients to be treated with pencil-beam-scanning proton radiotherapy.

Materials/Methods: A total of 47 proton therapy clinical plans generated by the treatment planning staff for esophageal cancer patients, previously treated at our clinic using the MEVION S250i pencil beam proton system, were used to construct a plan database, of which 40 plans were used to train and 7 plans to validate a 3D dense U-Net machine learning model for predicting 3D dose distributions. Inputs to the model include dose prescriptions, the target, the body contour, and four organ-at-risk (OAR) contours: the spinal cord, heart, lungs, and stomach. The model performance was assessed by comparing the DVH values of the predicted dose distributions with that of the clinical plans for 8 separate esophageal cancer treatment plans.

Results: Despite the variability in esophageal treatment plans, the predicted dose distributions in the test set exhibited comparable quality to the clinical plans. The average value of the mean absolute dose difference was 1.47 ± 2.68 Gy. The percentage dose differences in the predicted target metrics of D5% and D95% were 2.73% ± 1.42% and 12.36% ± 4.63%, respectively. For the OARs, the predicted mean and maximum dose differences were D_mean=2.48 ± 1.79 Gy and D_max =3.29 ± 2.77 Gy for the heart, and D_max=8.03 ± 4.98 Gy for the spinal cord.

Conclusion: We developed 3D DL models capable of rapidly predicting high-quality dose distributions for esophageal cancer proton therapy. The predicted dose distributions can be imported into a treatment planning system for generation of treatment plans with the use of scripting functions for beam placement and plan parameter optimization that produce the predicted dose distributions. The proposed DL-based planning approach can be used for decision-making before planning, individualized assessment of plan quality, and guiding automated planning to improve planning consistency and efficiency.