Scalable, memory-efficient robust proton therapy optimization through beamlet-free treatment planning

Proton therapy requires robust optimization to adequately take treatment uncertainties into account. Most robust optimizers utilize worst-case scenario selection methods, which significantly raises the computation time and memory requirements, since a beamlet matrix needs to be calculated for each scenario.

A beamlet-free robust optimization strategy that performs the spot weight optimization and scenario evaluation during the Monte Carlo dose calculation is investigated. This allows for reducing time and memory requirements for each scenario by omitting the costly beamlet matrix.

Three robust optimization methods, beamlet-based worst-case scenario, beamlet-based expected-value, and beamlet-free expected-value are presented and compared on six patient cases, regarding achieved plan quality and required computational burden.

Comparable plan quality could be achieved for all methods. Computational requirements between beamlet-based worst-case and beamlet-based expected value implementations were comparable. Beamlet-free robust optimization successfully reduced memory usage by $87\%$ to $92\%$ and computation time by $51\%$ to $69\%$ of the what is required by beamlet-based methods.

Beamlet-free robust optimization methods can maintain plan quality while lowering computational requirements, allowing the use of more challenging optimization conditions, such as higher numbers of scenarios, finer voxel resolution and bigger number of spots. It seems a promising tool to open new perspectives for novel proton therapy techniques.