Breast radiation therapy fluence painting with multi-agent deep reinforcement learning

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence

Background:

• Electronic compensation (ECOMP) offers excellent dose conformity in breast radiation therapy.
• Manual fluence painting for ECOMP is time-consuming and inefficient for clinical workflows.

Purpose of the Study:

• To develop an automated treatment planning tool for breast radiation therapy using reinforcement learning (RL).
• To enable iterative editing of fluence maps guided by clinical objectives for efficient planning.

Main Methods:

• A retrospective study of 70 patients treated with ECOMP using 6MV tangential photon beams.
• Each fluence map pixel acted as an RL agent, using beam-eye-view dose profiles as input.
• RL agents modified fluence maps to optimize target coverage and dose homogeneity, with plans compared to clinical standards.

Main Results:

• The RL tool generated ECOMP plans in approximately 90 seconds.
• RL-generated plans demonstrated comparable isodose distribution and dosimetric endpoints to manually created clinical plans.
• Key dosimetric parameters like Breast PTV V95% and V105% were similar between RL and clinical plans.

Conclusions:

• The developed RL framework successfully automates the generation of breast ECOMP plans.
• The AI-driven approach yields clinically acceptable plan quality efficiently.
• This technology has the potential to streamline radiation therapy planning processes.