ESTRO 38 Abstract book

S1171 ESTRO 38

EP-2116 End-to-end dosimetry audits of Stereotactic Ablative Radiotherapy M. Shaw 1,2 , A. Alves 1 , C. Davey 1 , M. Geso 2 , F. Kadeer 1 , J. Lehmann 3,4 , J. Supple 1 , J. Lye 1 1 ARPANSA, Australian Clinical Dosimetry Service, Yallambie, Australia ; 2 RMIT University, School of Health and Biomedical Sciences, Melbourne, Australia ; 3 Calvary Mater Hospital, Radiation Oncology, Melbourne, Australia ; 4 RMIT University, School of Science, Melbourne, Australia Purpose or Objective Stereotactic Ablative Radiotherapy (SABR) refers to radiotherapy treatment deliveries which give a high dose of radiation to an extra-cranial tumour, with high geometric precision (1). The risk to patients is increased with SABR techniques due to the increased dose per fraction, and as such specialised planning, treatment and QA practices are needed to ensure patient safety (1). Independent on-site audits are recommended by current The Australian Clinical Dosimetry Service (ACDS) began SABR audits in March 2018. An end-to-end dosimetry audit was developed using a customised CIRS® humanoid thorax phantom. The audit planning cases include lung, spine and soft tissue targets, with delivery modality being at the discretion of the treating facility. Point dose measurements were taken at the centre of each target volume, and in the spinal cord using a PTW 60019 microDiamond. Gafchromic EBT3 Film measurements were taken and analysed with gamma criteria and distance-to- agreement between planned and measured dose profiles. Results To date, 59 SABR plans have been included in the audit. Figure 1 shows the local point dose variation for the target measurements and the global point dose variation for the spinal cord measurements. Large variations are seen in the vertebrae hard bone measurement point (average –3.6%, standard deviation 2.4%) and the lung measurement point (average -2.7%, standard deviation 2.4%). The measurements in the spine point in Figure 1 are with the detector calibrated in water. To account for measurement in bone, algorithm dependent Monte Carlo corrections are applied and the average improves to 0.0%, standard deviation 2.3%. SABR guidelines (1-3). Material and Methods

The results indicate both a good accuracy and good generality of the trained model. ML model seems to be able to predict when a plan solution is deliverable during the planning processes, reducing the number of DQA failures, the re-optimizations due to failing DQA and the number of plans being rejected during the physics plan check because of overmodulation issue. EP-2115 Semi-automated quality assurance of deformable registration in CT radiotherapy data T. Mcgrath 1 , Z. Lawrence 1 , R. Farhad Salih 1 , Y. Peters 1 , J. Rawling 1 , M. Wilson 2 , C. Piazzese 3 , S. Holloway 2 1 The University of Manchester, School of Physics and Astronomy, Manchester, United Kingdom ; 2 University College London, Department of Medical Physics & Biomedical Engineering, London, United Kingdom ; 3 Cardiff University, School of Engineering, Cardiff, United Kingdom Purpose or Objective A common issue in many medical physics studies is the lack of data available in order to successfully train machine learning algorithms to be able to validate registrations where ground truths are not available. Usually large datasets are required, especially for inference on more intricate structures. The aim of this work was to establish the viability of performing quality assurance on Deformable Image Registration (DIR) through the use of Convolutional Neural Networks (CNNs) in CT images. Material and Methods A total of 94 head and neck cancer patients’ 2D planning CT images converted to 3D Nifty images from The Cancer Imaging Archive (TCIA) were used in this study. Pairwise affine and deformable registrations were performed on this dataset using the open source NiftyReg software package, and the contours of the brain stem, spinal cord, and left & right parotid glands were propagated. A brief grid search was carried out to optimise the bending energy, a penalty term in the optimisation algorithm to avoid overfitting, and control point spacing of the registrations, to ensure a reasonable quality of the registrations. The results were evaluated with commonly used overlap and surface based similarity metrics that measure how well the propagated contour has been transformed compared to the ground truth. Examples of these metrics include the DICE score and the 95th percentile surface distance. Furthermore, a brief viability study was conducted using pre-existing architectures within NiftyNet, by segmenting structures within a 3D CT scan of a patient. Results Using NiftyReg, 2753 out of a possible 8742 registrations were obtained. The average values of the DICE scores between the ground truth segmentation of the reference image and the propagated contours of the executed registrations were in the range of 0.5 - 0.6. However, there were a significant number of outliers, revealing the abundance of poor registrations and further highlighting the necessity of quality assurance in DIR. For the viability study, the method that provided the best metric values involved training 2D U-Net on 2D slices of the image and for a segmentation of the spinal cord to be generated for each one. The concatenated 3D segmentation was compared to a ground truth contour provided by a clinician by means of the metrics mentioned previously, obtaining a DICE score of 0.92 and a 95th percentile distance of 1.12. Conclusion To develop the project, the 2D U-Net method could be tested on the other structures to determine the difference in segmentation quality between them. Ideally, a 3D segmentation method would be successfully employed, as the field of medical physics is progressing in that direction. Combined with the registration dataset produced, our research lays the groundwork for continued development of this project.

Film localisation accuracy within the phantom has been verified to <0.5 mm. The distance-to-agreement in the sagittal and coronal planes was analysed at the 50% isodose level. Figure 2 shows the results of a VMAT spine plan with maximum discrepancies of -0.08 mm (A-P) and 0.55 mm (L-R).