Novel models and techniques in optimization for cancer treatment using high dose rate brachytherapy

Abstract

In this thesis, we explore three research lines concerning the optimization of cancer treatment plans, with a focus on high-dose radiation therapy. First, we present a novel linear network-based optimization model for treating cancers with high-dose rate brachytherapy (HDR-BT). Our model measures the discrepancy between the dose delivered and the prescription in a novel way. It generalizes several existing models, such as Inverse Planning by Simulated Annealing (IPSA) which aims to minimize the deviation between dose delivered and prescription for each voxel individually. This model improves the distribution of hot spots (areas receiving too much dose) throughout the tumor tissues. Second, we propose three different optimization models for needle placement in 3D-printed masks for HDR-BT. The current alternative method employs a rigid template that help physicians position needles. However, such templates can restrict the relative positions of needles, which may result in lower quality treatment plans. The models we propose take advantage of designing 3D printing masks, allowing more freedom in needle placement and yielding plans with better dosimetric indices. Third, in optimization models for HDR-BT radiation treatment planning, simplification assumptions are often made to avoid complicated dose computations and to create models that can be described in closed-form. To incorporate more precise dose computations, these models may need to be formulated as black-box optimization problems, for which model components are not known in closed-form but can be evaluated at selected points. In this context, we explore the computational potential of a new surrogate model that is inspired from the way tree ensembles are constructed for predicting constraint and objective function values from observed data. Unlike traditional tree-based models, this surrogate does not require training of trees and is faster to optimize.