ESTRO 2022 - Abstract Book

S470

Abstract book

ESTRO 2022

The developed novel 4pi-DTRT path-finding is an alternative to geometry based path-finding for DTRT treatment planning, resulting in similar plan quality for common HN cases. It has the potential, however, of increased efficiency through automation. This work was supported by Varian Medical Systems, Inc.

MO-0545 A quasi-optimal non-coplanar 4 π -VMAT solution for treating head & neck cancers

J. Simms 1 , C. Rowbottom 2 , R. Dawson 2

1 The Clatterbridge Cancer Centre, Radiotherapy Physics , Liverpool, United Kingdom; 2 The Clatterbridge Cancer Centre, Radiotherapy Physics, Liverpool, United Kingdom Purpose or Objective Patients treated with radiotherapy for H&N cancer unfortunately suffer from high rates of post treatment complications, due to the significant number of radiosensitive OARs within close proximity to large tumour volumes. Non-coplanar 4 π - VMAT with increased degrees of freedom has the potential to create more conformal plans that would likely benefit this patient group. The aim of this study was to develop and independently test a quasi-optimal non-coplanar 4 π -VMAT configuration. Materials and Methods The study was divided into two main phases. In phase one, a cohort of ten previously treated H&N patients were replanned using a 12-arc plan; 11 non-coplanar arcs and a coplanar arc. The 12-arc plans were reduced to a single arc plan by using an iterative algorithm based on MU. The arc contributing the fewest MU was deleted from the plan, the plan re-optimised and the process repeated. Consequently, it was possible to determine how the number of arcs in a plan affects certain dose metrics and which couch rotation angles seem preferable for this patient group. In phase 2, a preferred non-coplanar 4 π - VMAT configuration developed from phase 1 was applied to an independent cohort of ten H&N patients, and the effect on OAR doses assessed. The resulting plans were delivered and timed for comparison with the standard 2-arc coplanar approach. Results The gradient index, along with spinal canal and parotid doses, were minimised when as few as three non-coplanar arcs were used. In addition, non-coplanar arcs with large couch rotation angles survived for longer during the iterative process and therefore were deemed superior. In phase two, statistically significant dose reductions were obtained when using a 4- arc solution consisting of a coplanar arc and three non-coplanar arcs with 90° and ±75° couch rotations. These included mean reductions in the doses to the spinal canal D 0.1cc (3.83±1.33 Gy, p < 0.001), brainstem D 0.1cc (5.15±3.71 Gy, p = 0.019), contra-lateral parotid D mean (3.83±2.06 Gy, p = 0.005) and ipsi-lateral parotid D mean (4.18±1.99 Gy, p = 0.002). The mean delivery time was 450±32 s, compared to 170±9 s for the original, 2-arc coplanar plans. The 4 π -VMAT delivery time can likely be reduced via intelligent sequencing of arcs to minimise gantry/couch rotations and applying single button press delivery techniques. Assessing the dosimetric accuracy of these plans will form part of the ongoing work. Conclusion By using a simple, iterative algorithm, a quasi-optimal non-coplanar class-solution was developed from a small cohort of patients. Non-coplanar VMAT for H&N cancer shows promise in reducing OAR doses and therefore side effects from treatment. M. Lempart 1,2 , C. Jamtheim Gustafsson 1,2 , M. Nilsson 3 , M. P. Nilsson 1 , N. Svanberg 1 , J. Scherman 1 , P. Munck af Rosenschöld 1,4 , L. E. Olsson 1,2 1 Skåne University Hospital, Department of Hematology, Oncology, and Radiation Physics, Lund, Sweden; 2 Lund University, Department of Translational Sciences, Medical Radiation Physics, Malmö, Sweden; 3 Lund University, Centre for Mathematical Sciences, Lund, Sweden; 4 Lund University, Department of Medical Radiation Physics, Lund, Sweden Purpose or Objective Treatment planning for external beam radiation therapy (ERBT) involves the manual segmentation of target and organs at risk (OAR) structures on computed tomography (CT) images. Deep learning technologies can automate this task. Another imaging modality that can be used for segmentation is cone beam computed tomography (CBCT), usually used for patient alignment, but advantageous in adaptive radiation therapy. Training a deep learning segmentation model often requires large datasets with ground truth (GT) delineations. For the CBCT domain, such GT data is limited, due to today’s clinical routines. The aim of this work was therefore to develop a deep learning model trained on CT images, that can be used for both, CT and CBCT segmentation of prostate target and OAR structures. To make the model compatible with CBCT images, we propose a stacked image augmentation pipeline, combining CBCT, domain specific and standard augmentations techniques. Materials and Methods A deeply supervised 3D U-Net was trained on 342 CT volumes using standard augmentations (rotation, scaling and mirroring), serving as the baseline model. An improved model using the same architecture was then trained with standard augmentations, as well as additional augmentations (elastic deformation, gamma, additive brightness) and CBCT, domain specific augmentations (intensity shift, speckle noise, randomized piece-wise linear histogram matching). Both models were trained using three-fold cross validation and final models were established by model ensembling. For histogram matching, a separate dataset of 50 CBCT image volumes was used. With a probability of p = 0.3, the histogram of a CT volume was matched against the histogram of a randomly selected CBCT during model training. Trained models were applied to a test MO-0546 Data augmentation for deep learning multipurpose prostate target and OAR segmentation in CT and CBCT