Click here to

Session: Multi-Disciplinary: Innovative Technologies [Return to Session]

Development of a Deep Learning Model That Can Create Multiple Auto-Plans Prioritizing Different Clinical Goals to Facilitate Decision Making in Complex Pancreatic VMAT with Dose Painting

Y Wang*, J Wo, T Hong, Department of Radiation Oncology Massachusetts General Hospital, Harvard Medical School, Boston, MA

Presentations

TU-IePD-TRACK 3-7 (Tuesday, 7/27/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: To develop a deep learning (DL) model that can create multiple auto-plans prioritizing different clinical goals to support decision making in complex pancreatic VMAT with dose painting.

Methods: A DL model was trained with 109 VMAT plans in RayStation 10B. The pancreatic tumor and regional nodes receive 50.4 Gy, with the involved vessels painted to 58.8 Gy. The model used three different sets of dose prediction and mimicking rules to generate auto-plans prioritizing PTV coverage, OAR dose and the balance between the two. The model was validated and tested on ten and 20 new patients, respectively. Each auto-plan was compared to the clinical plan using 12 clinical goals, with the p value calculated by a two-tailed t-test. The auto-plan most similar to the clinical plan was selected as the favorable plan.

Results: On average, all three auto-plans offered some universal advantages over the clinical plan: slightly higher CTV coverage, significant reduction in stomach mean dose, kidney V18 and V10, and V50 for GI organs excluding PTV5040. For PTV coverage, uniformity in low-dose PTV and GI maximum dose, the OAR auto-plans were the most similar to the clinical plans, whereas the PTV auto-plans were the least. The PTV, balance and OAR auto-plans were picked 20%, 40% and 40% of the time as the favorable plan, respectively. Collectively, the favorable auto-plans offered similar performance to the clinical plans on seven goals and significant improvement on five. 30% of the favorable auto-plans were directly acceptable. 20% and 40% required minor and moderate processing, respectively. 10% exhibited large difference, but due to significant relaxation of some OAR constraints in the clinical plan.

Conclusion: For patients with standard OAR constraints, the DL model could always produce an auto-plan very similar to the clinical plan, which was acceptable directly or after minor or moderate processing.

ePosters

    Keywords

    Treatment Planning

    Taxonomy

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

    Contact Email

    Share: