Purpose: To evaluate the ability of our machine learning model to predict 3D dose distributions that can be converted to deliverable plans using the same and alternate treatment techniques than the training set plans.
Methods: We developed a triple-stage cascaded U-Net model for 3D dose prediction for head and neck treatments using the data provided for the 2020 AAPM OpenKBP planning challenge. These plans were created using nine field step-and-shoot IMRT, prescribed to deliver 70 Gy in 35 fractions to a high risk PTV, and potentially 63 Gy to an intermediate risk PTV and/or 56 Gy to a standard risk PTV. We used our model to predict dose for 11 patients and imported the data into Raystation. Using the fall-back planning function, we created two plans that attempted to mimic the predicted dose, one using the same 9 field IMRT beam arrangement as the training plans and one using VMAT with two full arcs, which is the standard at our institution. The plans were normalized to give 95% coverage at 70 Gy to the high risk PTV. These plans were compared against the predicted dose based on several main clinical objectives: coverage of the standard risk PTV, max dose to the cord, mean dose and D50 to the left and right parotids, and max dose in the plan.
Results: There were no statistically significant differences in the majority of the clinical goals (p>0.05), with the exception of the max spinal cord (for IMRT and VMAT) and the max overall (IMRT) dose. Adding an optimization objective for these goals in addition to the mimic dose objective for select plans removed this difference.
Conclusion: Our model generates 3D dose predictions that are clinically achievable both using the technique used in the training plans and an alternate technique.
Treatment Planning, Radiation Therapy
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation