Search

Automated radiotherapy planning is characterised by reduction in manual planning due to an increase in computerised planning. Current methods can produce plans suitable for clinical use. However, every case is unique and manual intervention is often needed. The goal of this work was to determine whether it is feasible to develop a fully automated planning system producing clinically optimal plans, and if so, to begin developing it. This work explored relationships between automated planning parameters and anatomical features with respect to dosimetric outcomes. A rules-based automated planning technique was used, an algorithm requiring calibration of input parameters prior to use. This calibration determines the target objectives the algorithm will optimise to. Existing calibration methods use a single set of calibrated parameters per treatment site and are applied to all patients. This approach is considered sufficient to meet clinical goals but may not be sufficient for development of optimal personalised planning due to anatomical variance between patients. Using a validated rules-based planning methodology and obtaining patient bespoke expert-driven calibrated parameters as the optimal gold standard and validation benchmark, two machine learning techniques were explored for apriori configuration of parameters for the delivery of personalised treatment planning. The main objective was to train models to predict gold standard parameters hence generating expert planning automatically. A secondary objective was to determine dosimetric differences between plans generated via machine learned parameters and a traditional single set of parameters applied to all cases. Preliminary studies were carried out to define what will be considered gold standard and to identify anatomical features for inclusion in the main study as well as their relationships to calibrated parameters. The research presented here was applied to three sites: prostate, rectum and lung. Findings are a

Automated radiotherapy planning is characterised by reduction in manual planning due to an increase in computerised planning. Current methods can produce plans suitable for clinical use. However, every case is unique and manual intervention is often needed. The goal of this work was to determine whether it is feasible to develop a fully automated planning system producing clinically optimal plans, and if so, to begin developing it. This work explored relationships between automated planning parameters and anatomical features with respect to dosimetric outcomes. A rules-based automated planning technique was used, an algorithm requiring calibration of input parameters prior to use. This calibration determines the target objectives the algorithm will optimise to. Existing calibration methods use a single set of calibrated parameters per treatment site and are applied to all patients. This approach is considered sufficient to meet clinical goals but may not be sufficient for development of optimal personalised planning due to anatomical variance between patients. Using a validated rules-based planning methodology and obtaining patient bespoke expert-driven calibrated parameters as the optimal gold standard and validation benchmark, two machine learning techniques were explored for apriori configuration of parameters for the delivery of personalised treatment planning. The main objective was to train models to predict gold standard parameters hence generating expert planning automatically. A secondary objective was to determine dosimetric differences between plans generated via machine learned parameters and a traditional single set of parameters applied to all cases. Preliminary studies were carried out to define what will be considered gold standard and to identify anatomical features for inclusion in the main study as well as their relationships to calibrated parameters. The research presented here was applied to three sites: prostate, rectum and lung. Findings are a