Evolutionary Brachytherapy Treatment Plan Optimisation using Low-Resolution Dose Calculations

Abstract

Radiotherapy treats cancer by irradiating tumour cells so that they stop dividing and die off. High-dose-rate (HDR) brachytherapy is a form of radiotherapy that delivers radiation from inside the body, allowing a high dose in the tumour while sparing surrounding tissue. The effectiveness of HDR brachytherapy depends on how radiation is distributed in a patient’s body which is determined by a treatment plan. HDR brachytherapy treatment planning for cervical cancer can be automated using evolutionary algorithms (EAs), as is done by the BRIGHT treatment planning toolkit, which uses the EA MO-RV-GOMEA. Treatment plans found by BRIGHT, after several minutes of optimisation using a graphics processing unit (GPU), have been shown to match or outperform current clinical practice. Even faster treatment planning could lower hardware requirements, shorten the time a patient spends with the applicator in place, or yield better treatment plans in the same optimisation time. Treatment planning using BRIGHT requires many evaluations of candidate plans, each computing the dose at many points, which dominate optimisation time. Using a lower dose calculation resolution can enable faster treatment planning, yet allows the EA to overfit to the resulting imprecise approximation, which risks producing worse treatment plans. This thesis studies whether and how treatment planning using BRIGHT can be made faster, without loss of plan quality, by using a lower dose calculation resolution.

MO-RV-GOMEA is first studied in benchmark problems with artificial uncertainty. Later, two modifications to BRIGHT for reducing overfitting are proposed. Using a lower dose calculation resolution is shown to yield better plans than the baseline during the first 10 seconds of optimisation, but never improves on the final plan quality in terms of sparing, coverage and shape metrics. Independent of the dose calculation resolution, rounding errors stemming from partial evaluations are shown to negatively impact optimisation more than previously anticipated, and using full evaluations is shown to be at least as fast as partial evaluations in the GPU-parallelised version of BRIGHT.