Deep learning-based inverse mapping for fluence map prediction.

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Healthcare

Background:

• Volumetric Modulated Arc Therapy (VMAT) planning involves complex optimization processes.
• Accurate fluence map generation is crucial for effective VMAT treatment delivery.
• Current optimization methods can be time-consuming, impacting clinical workflow.

Purpose of the Study:

• To develop and evaluate a novel, optimization-free method for predicting fluence maps in VMAT planning.
• To enable direct generation of fluence maps from desired dose distributions.
• To significantly reduce the time required for VMAT treatment planning.

Main Methods:

• A two-step prediction method involving deep neural network inverse mapping from dose projections to phantom fluence maps.
• Application of a plan scaling technique to translate phantom fluence maps to patient geometry.
• Evaluation using 102 head and neck and 14 prostate cancer VMAT plans.

Main Results:

• High accuracy in predicting fluence maps, with mean dose differences of 1.42% ± 0.37% for PTV.
• Excellent dose agreement, evidenced by a 98.06% ± 2.64% gamma passing rate (3 mm/3% criterion).
• Prediction time for a single VMAT plan was less than one second.

Conclusions:

• The developed inverse mapping-based method accurately predicts fluence maps for desired dose distributions in VMAT.
• This optimization-free approach is orders of magnitude faster than traditional fluence map optimization.
• Integration with leaf sequencing holds potential for dramatically accelerating VMAT treatment planning.