ESTRO 2020 Abstract book

S173 ESTRO 2020

implemented for the first time in the world. Initial experience demonstrates online adaptive sessions to be achievable within our standard time slots and furthermore indicate a potential in reducing the treated volume and thus also toxicity related dose parameters. The work to prove the latter by clinical studies is ongoing. PD-0309 Comparison of CBCT based synthetic CT methods for adaptive proton therapy A. Thummerer 1 , P. Zaffino 2 , A. Meijers 1 , G. Guterres Marmitt 1 , J. Seco 3 , R.J.H.M. Steenbakkers 1 , J.A. Langendijk 1 , S. Both 1 , M.F. Spadea 2 , A.C. Knopf 1 1 University Medical Center Groningen- University of Groningen, Department of Radiation Oncology, Groningen, The Netherlands ; 2 Magna Graecia University, Department of Experimental and Clinical Medicine, Catanzaro, Italy ; 3 Deutsches Krebsforschungszentrum DKFZ, Department of Biomedical Physics in Radiation Oncology, Heidelberg, Germany Purpose or Objective Adaptive proton therapy (APT) aims to reduce dose to organs at risk (OAR) and ensures target dose coverage by frequently adapting treatment plans to anatomical changes. Repeated imaging, such as cone-beam computed tomography (CBCT), which in radiotherapy is commonly used for patient alignment purposes, also facilitates plan adaptation decisions. However, CBCT images suffer several image quality deficiencies (e.g. scatter) that prevent direct proton dose calculations. To overcome this limitation, several techniques to correct CT-numbers of CBCTs and to subsequently allow accurate dose calculations have been proposed in literature. In this study we compared three of these methods using a large head and neck dataset. Resulting synthetic CTs were not only evaluated in regard to their image quality and dosimetric accuracy, but also in terms of clinical suitability (e.g. conversion time) of each method. Material and Methods The comparison included a deep convolutional neural network (DCNN), deformable image registration (DIR) and an analytical image-based correction (AIC) method. Evaluation was performed on a dataset comprising 33 head and neck cancer patients treated with pencil-beam scanning proton therapy. For each patient a planning CT (pCT), weekly repeated CTs (rCT) and daily CBCTs were available. Image quality of synthetic CTs was quantified by calculating mean absolute error (MAE), mean error (ME) and the Dice similarity coefficient (DSC) of bone. Dosimetric characteristics were determined by intracranial proton dose calculations. Gamma pass ratios (2%/2mm and 3%/3mm) and relative range shifts were calculated to quantify dosimetric differences between the various sCT methods. Results On average, the lowest MAE and the highest DSC were observed for DCNN based sCTs (37 HU/0.96). Using DIR resulted in an average MAE of 44 HU and a DSC of 0.94. The highest MAE and the lowest DSC were observed for AIC based synthetic CTs, with values of 77 HU and 0.90 respectively. Figure 1 presents difference images between the various synthetic CTs and the reference rCT. The increased MAE for AIC is clearly visible. For a 2%/2mm passing criteria, gamma analysis resulted in average pass ratios of 99.30 %, 98.65 % and 97.35 % for DCNN, DIR and AIC respectively. A similar trend was observed for relative range shifts. Figure 2 shows a dose and CT-number profile comparison. The time to create a synthetic CT was 1 minute for AIC, 3 minutes for the DCNN, and 20 minutes

Conclusion Overall a strong correlation was observed between motion at planning and first treatment. Patients with poor 4D-CT image quality could be closely followed at first treatment to verify the motion. For motion differences larger than 0.6 cm either replanning with different margins or a resimulation may be warranted. PD-0308 First clinical experience of online adaptive radiotherapy driven by CBCT and artificial intelligence P. Sibolt 1 , L. Andersson 1 , L. Calmels 1 , C.F. Behrens 1 , E. Serup-Hansen 1 , H. Lindberg 1 , L. Sonne Mouritsen 1 , D. Sjöström 1 , P. Geertsen 1 1 Herlev & Gentofte Hospital, Radiotherapy Research Unit- Department of Oncology, Herlev, Denmark Purpose or Objective As a result of the rapid development of dynamic radiotherapy over the last decade it is today possible to create advanced treatment plans, sparing organs at risk while providing conform dose coverage of the target volume. However, one major factor limiting the reduction of the treatment volume is margins needed to account for inter- and intra-fractional anatomical variations. The purpose of this project is to describe the first clinical implementation of a commercial solution for CBCT-guided online adaptive radiotherapy (oART), driven by artificial intelligence (AI), enabling elimination of margins otherwise necessary to account for inter-fractional variations. Material and Methods As the first clinic in the world we have implemented a new commercial solution for AI-driven CBCT-based daily oART. The system applies structure-guided deformation of targets and organs at risk, from original definition on a reference CT to the high quality CBCT, based on initial AI driven auto-segmentation of so-called influencer organs. Automated treatment planning and calculation-based QA then enables the choice of a re-optimized plan on the anatomy of the day. Simulated online adaptive sessions were conducted by extensive use of an emulator, which laid a foundation for commissioning and clinical implementation, enabling the first treatments in the world on this new system. The system was systematically evaluated in terms of speed, plan quality, structure propagation accuracy, and treatment delivery. Results Emulator work during commissioning rendered in >600 automatic treatment plans and >100 systematically simulated online adaptive sessions, with the majority of the online sessions resulting in none or only minor editing of automatically generated structures and the choice of plan being the adapted in >90% of the treatment sessions. Limitations in auto-segmentation is currently correlated to cases where the system is not yet trained, e.g. urinary catheter, and where segmentation is also challenging for the human eye, e.g. seminal vesicles. The online adaptive process was for all the five first treated patients completed within 15-20 minutes, with influencer segmentation within 3-5 minutes and target definition within an additional 1-3 minutes. Bladder patients treated had a 40% average reduction in treatment volume, e.g. resulting in up to 30% reduction in V 45Gy and V 30Gy to the bowel bag, compared to the standard non-adaptive

treatments. Conclusion

A novel commercial solution for CBCT-based oART has been demonstrated to deliver accurate and fast adaptation to the anatomy of the day and was clinically