Room: Exhibit Hall | Forum 7
Purpose: To automate the beam orientation selection process in radiation therapy via policy iteration. The resulting policy should produce a near real-time beams selection for patients.
Methods: A tree lookout strategy searches for beams appropriate for a patientâ€™s CT map based on an L2-normed fluence map objective (FMO) and a heuristic â€œangles collisionâ€? measure during MDP episodes until a desired beam plan is achieved. The â€œangles collisionâ€? heuristic ensures beams spread within a plan while the FMO objective produces a distribution over global beam-space. Episodic trajectories are stored along nodes and edges of a tree. Multiple search episodes traverse different tree branches whilst progressively averaging the optimal FMO per node. A greedy tree policy based on the running FMO distribution recommends beam selection at the following trajectory iteration. Over multiple episodes, multiple branches for different patients are added â€“ producing a diverse FMO distribution for various beam-patient combinations. After search, the search data is parameterized by a large function approximator that maps patient geometries to beam angles. This approximator efficiently selects beams for test cases in a real-time scenario.
Results: The DVH and dose distributions are compared against equiangular and column generation derived-plans on prostate cancer cases. Our scheme produces plans with lower fluence costs, and lower dose delivered to critical structures compared to these other schemes.
Conclusion: A single-agent rolls out multiple searches, generating a distribution over possible beam orientations for every patient in a training dataset. A large deep neural network policy then learns the mapping from patient geometries to beams. Results reflect lower beam selection times, and improved performance compared to equiangular and column generation schemes. This can reduce the beams selection search time and improve plan quality.
Not Applicable / None Entered.