Linda Rossi

In conventional, manual planning for IMRT, a planner has to make many choices to drive the treatment planning system (TPS) towards generation of a high-quality plan. In an interactive trial-and-error procedure, he/she has to carefully select the number of beams, the beam directions, as well as the cost functions with their weights to define an optimization problem that will result in a high-quality plan. Each problem definition results in a treatment plan with unique trade-offs between all treatment objectives, and definition of the optimization problem that results in optimal trade-offs is highly complex. As a consequence, the quality of a radiotherapy treatment plan may be highly dependent on the skills of the involved planner and on allotted planning time. This is the case in a single plan generation for a patient, but it can also affect research studies where plan comparisons are performed to compare treatment techniques.

Automated planning with our in-house Erasmus-iCycle optimizer has demonstrated possibilities for consistent generation of high-quality plans, while fully avoiding manual planning workload. As a unique feature, this system features integrated beam profile and beam angle optimization (BAO).

In this thesis, Erasmus-iCycle was used to systematically investigate the impact of beam configurations on plan quality, and to investigate plan quality improvements relative to conventional manual planning.

In Chapter 2, the BAO option was used to investigate relationships between plan quality and the beam angle search space, i.e. the set of candidate beam directions that may be selected for generating an optimal plan. Ten prostate SBRT patients were included in the study. Autoplans with up to 30 beams with individualized directions were generated for 5 different candidate beam sets, one coplanar and four non-coplanar. The candidate sets were: i) a coplanar set, covering the whole 360° range (CP), ii) all directions (mainly anterior) available in the robotic CyberKnife treatment unit (CK), iii) a fully non-coplanar sphere (F-NCP), i.e. also including posterior beams that were not present in CK, and iv)/v) CK+ and CK++, as subsets of F-NCP, with higher beam density than the CK beam set (CK+), or covering a (bigger) laterally extended beam area (CK++). In total 1500 plans were generated. Generated plans were clinically acceptable, according to an assessment of involved clinicians. All plans were generated with highly similar PTV coverages, allowing plan comparisons to be based on OAR dose parameters, with the rectum considered most important. OAR sparing improved with all NCP configurations compared to CP, especially for the rectum. F-NCP performed the best, with reductions in rectum Dmean, V40GyEq, V60GyEq and D2% of 25%, 35%, 37%, and 8%, respectively, compared to CP. CK performed slightly worse than F-NCP, which could be compensated by the laterally extended beam area in CK++. Addition of posterior beams (CK++ -> F-NCP) or enhancement of the beam density (CK -> CK+) did not lead to further improvements. Increasing the number of selected beams significantly improved plan quality. For coplanar plans, for instance, rectum Dmean, V40GyEq, V60GyEq and D2% could be improved by 39%, 57%, 64%, and 13%, respectively, when using 25 beams instead of 11 beams. Using more than 25 beams did not result in relevant further plan improvements.

With the clear benefit of non-coplanar Cyberknife beams for prostate SBRT as observed in Chapter 3, the possibility of creating a non-coplanar beam angle class solution (CS) for Cyberknife was explored to replace the time-consuming individualized BAO, while not losing in plan quality. CS generation was performed in 3 steps, based on 10 training patients. First, Erasmus-iCycle was used to generate plans with 15, 20, and 25 non-coplanar individualized beams. Secondly, based on the beams selected in these plans, 6 recipes for creation of beam angle CSs were investigated for all three beam numbers. Finally, Erasmus-iCycle was used to generate plans for the fixed (6x3) CSs, both for the 10 training patients and for 20 independent validation patients. A total of 1060 plans was generated. Out of the 6 tested CS recipes, only 1 resulted in 15-, 20-, and 25-beam non-coplanar CSs without plan deterioration compared with individualized BAO. Negligible differences were found between 25-beam CS plans and 25-beam BAO plans, with mean differences in rectum D1cc, V60GyEq, V40GyEq, and Dmean of 0.2 ± 0.4 Gy, 0.1 ± 0.2%-points, 0.2 ± 0.3%-points, and 0.1 ± 0.2 Gy, respectively. Differences between 15- and 20-beam CS and BAO plans were also negligible. On the other hand, computation times with the CSs were reduced by a factor of 14 to 25, due to the avoidance of costly BAO.

In Chapter 4, the first system for fully automated generation of clinically deliverable CyberKnife treatment plans (autoROBOT) was developed and evaluated for prostate SBRT. To this purpose, Erasmus-iCycle was coupled to the commercial CyberKnife TPS. The system was first validated by comparing automatically generated CyberKnife plans with manually generated plans. Next, for 20 patients, autoROBOT plans were compared to VMAT plans, that were also automatically generated (autoVMAT). Both autoROBOT and autoVMAT plans with CTV-PTV margins of 3 mm (as used in clinical prostate SBRT CyberKnife routine) were generated. In addition, 5 mm CTV-PTV margin autoVMAT plans were generated (a margin often applied for VMAT). Compared to manual planning, autoROBOT improved rectum D1cc (16%), V60GyEq (75%) and Dmean (41%), and bladder Dmean (37%) (all p <= 0.002), with equal PTV coverage. Compared to autoVMAT with equal 3 mm margin, autoROBOT reduced rectum D1cc by 5% (p = 0.002), rectum V60GyEq by 33% (p = 0.001), and rectum Dmean by 4% (p = 0.05), respectively, with comparable PTV coverage and other OAR sparing. For autoVMAT with 5 mm margin, 18/20 plans had a PTV coverage lower than requested (<95%) and all plans had higher rectum doses than autoROBOT (mean percentage differences of 13%, 69% and 32% for D1cc, V60GyEq, and Dmean, respectively (all p < 0.001)). In Chapter 5, a similar workflow was developed for automated planning in robotic CyberKnife radiosurgery of benign vestibular schwannoma tumors, to explore possibilities for reducing dose outside the PTV to potentially reduce risk of secondary tumor induction. The goal of automated planning was to reduce the dose bath, including the occurrence of high dose spikes leaking from the PTV into normal tissues, without worsening PTV coverage, OAR doses, or treatment time. CyberKnife autoplans were generated for 20 patients, treated with 1x12 Gy, and compared with manually generated CyberKnife plans. Autoplans performed as good as manual plans for all OAR sparing (largest mean difference for all OARs: ΔD2% = 0.2 Gy), while highly reducing the dose bath. With autoplans, patient volumes receiving more than 1 or 6 Gy, were reduced by (mean/maximum reduction) 23.6/53.8% and 9.6/28.5% with autoplans compared to manual plans (p < 0.001). Autoplans also reduced dose spikes, with mean/maximum reductions of 22.8/37.2% and 14.2/40.4% in D2% for shells at 1 and 7 cm distance from the PTV, respectively (p < 0.001). The study showed that automated planning highly outperformed manual planning, reducing 'for-free' the dose bath outside the PTV, without deteriorating PTV coverage or OAR sparing, or significantly increasing treatment time. In Chapter 6, Erasmus-iCycle was challenged with planning for young female mediastinal lymphoma patients with large variations in tumor location, shape and size. The purpose of this work was to implement an automated planning workflow to obtain adequate target coverage with maximum sparing of breasts, heart, and lungs, and to investigate the impact of beam configuration on plan quality. Twenty-four coplanar and non-coplanar beam configuration approaches were considered, partly based on individualized beam angle optimization, and partly on beam angle class solutions. Twenty-six patients were included in the study. The automated planning workflow was first validated by comparing clinically delivered, manually generated plans (CLIN) with automatically generated plans. Next, for the beam configuration investigations, autoplans were generated with i) coplanar configurations with computer-optimized patient-specific beam directions (CP_x with x = 5-15), ii) non-coplanar configurations with patient-specific beam directions (NCP_x with x = 5-15), iii) the VMAT coplanar beam angle class solution, and iv) the non-coplanar Butterfly VMAT (B-VMAT) class solution. Of the 645 generated autoplans, 98.8% were suited for clinical use. Compared to the CLIN plans, autoplans had significantly enhanced PTV dose delivery and, especially for non-coplanar autoplans with 10-15 individualized beams, also large OAR dose reductions could be obtained. None of the investigated 24 beam configuration approaches was best for all patients, but overall non-coplanar configurations (B-VMAT and NCP_x >= 12) performed clearly the best. NCP_x >= 12 produced on average highly conformal plans with favourable high dose plan parameters for the lungs and the patient, and also a low heart Dmean. B-VMAT had reduced low-dose spread in lungs and left breast, with the practical advantages of a faster delivery and the elimination of patient-specific BAO. Generation of multiple plans for each new patient for a per-patient selection of the optimal beam configuration, based on both plan quality differences and practical considerations as delivery time, could importantly contribute to personalization of the treatment of these patients.

In Chapter 7, autoplanning with Erasmus-iCycle was used to explore the use of VMAT+, i.e. coplanar VMAT supplemented with a few (<= 5) non-coplanar beams, for enhancing OAR sparing in prostate SBRT with minimal increase in treatment time compared to VMAT. The work was inspired by successes reported for VMAT+ in liver SBRT [157]. Initially, VMATHS plans, complementing VMAT with five non-coplanar IMRT beams with computer-optimized, patient-specific directions, were generated for the 20 study patients, showing large preferences for a few principal directions in the beam angle search space. Two most preferred directions were used to define a 2-beam non-coplanar beam angle class solution (CS) for complementing VMAT, resulting in the VMATHCS treatment approach. VMATHCS autoplans were then compared to i) VMAT, ii) VMATHS, and iii) IMRT with 30 individualized non-coplanar beam directions (30-NCP). Plan comparisons were performed in terms of PTV dose, OAR sparing, and computation and treatment delivery times. Compared to VMAT, plan quality was significantly improved with the non-coplanar VMATHCS. For equal PTV dose, rectum Dmean, D1cc, V60GyEq and V40GyEq were reduced by 19.4 ± 10.6%, 4.2 ± 2.7%, 39.7 ± 23.2% and 34.9 ± 0.3%, respectively (all p < 0.001). Total delivery times only increased by 1.9 ± 0.7 min compared to VMAT (9.1 ± 0.7 min). VMATHCS performed equivalently to VMATHS regarding plan quality, while reducing optimization times by a factor of 25 due to avoidance of BAO. VMATHCS had larger dose bath than 30-NCP, but with equal quality regarding all other plan parameters and with highly reduced optimization and delivery times. In Chapter 8, challenges and opportunities of autoplanning with Erasmus-iCycle, and of the use of non-coplanar beam configurations are discussed. The chapter concludes with an outlook on future research opportunities.