A Novel Machine Learning Model for Dose Prediction in Prostate Volumetric Modulated Arc Therapy Using Output...

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning

Background:

• Prostate volumetric modulated arc therapy (VMAT) treatment planning is time-consuming, limiting exploration of treatment options.
• Current planning involves 5-30 minutes for optimization and calculation per plan.

Purpose of the Study:

• To develop a machine learning model to predict dose distributions for prostate VMAT, bypassing time-intensive optimization and calculation.
• To enable faster estimation of feasible dose objectives for multi-criteria optimization (MCO).

Main Methods:

• A novel voxel-wise residual network model was developed to predict dose distributions.
• The model utilizes optimization priorities and dose map shapes for initialization and refinement.
• Contiguous and atrous patch sampling were employed to enhance receptive fields and model efficiency.

Main Results:

• The model achieved modest average dose map root-mean-square errors (RMSEs) of 2.38 ± 0.47% of prescription dose.
• Cross-validation on 100 prostate VMAT cases demonstrated model accuracy.
• Optimal performance was observed with training set sizes between 60 and 90 patients.

Conclusions:

• The developed machine learning model can significantly accelerate prostate VMAT treatment planning.
• Faster dose distribution estimation allows for more time for plan refinement, potentially improving plan quality.
• This approach facilitates rapid exploration of the Pareto set of dose objectives.