ESTRO 38 Abstract book

S4 ESTRO 38

proton beam with an adaptive aperture for pre-clinical research. Br J Radiol 91: 20180446, 2018

Abstract text Artificial intelligence is increasingly being proposed for radiotherapy. It has the potential to support or even take over several factors in the radiotherapy chain. A few examples are automated delineation, prediction of dose calculation, guiding optimal plan optimization, QA of treatment plans, and verification on the linac. Generally, AI is based on historic information, and a mathematical model, or a deep learning convolutional neural network is built on the basis of a library of data which is used for training. The result is a product that can take over clinical tasks from clinicians, dosimetrists, physicists and technologists. Therefore, such product is considered as a medical device and it should obey the EU rules for medical devices. The data that is used for training is in general patient related data, and the hospital should follow the privacy rules / get patient consent / get permission from the medical ethics committee / use only anonymized data. Furthermore, the clinical data may not be as consistent as we hope. For contouring, it may contain wrongly assigned structure names, or quickly contoured structures that deviate from the standard, and for dose prediction, it may contain sub-optimal treatment plans where insufficient OAR sparing was achieved. This can influence the performance of the AI. The AI result will not necessarily be worse than the contour by the clinician, but it is advisable to always check the output of an AI system. Ideally, a prediction would include an uncertainty with which the prediction is made, which is an important field of work in deep learning. Finally, the output of an AI algorithm also needs to be accepted by the medical staff, which requires convincing testing. SP-0011 Unified Radiogenomic Prediction of Late Radiotherapy Toxicities J. Coates 1,2 , A. Jeyaseelan 2 , N. Ybarra 2 , J. Tao 2 , M. David 3 , S. Faria 3 , L. Souhami 3 , F. Cury 3 , M. Duclos 3 , I. EL Naqa 4 1 Department of Oncology, University of Oxford, Oxford, United Kingdom ; 2 Mcgill University, Medical Physics Unit, Montreal, Canada ; 3 Mcgill University, Radiation Oncology Division, Montreal, CANADA; 4 University of MIchigan, Department of Radiation Oncology, Ann Arbor, USA Abstract text Radiation-induced toxicities are the unavoidable result of damage to healthy tissues that surround solid tumours and directly limit the ability of clinicians to achieve local tumour control. Clinically, they present as acute events, becoming manifest and subsiding within days-to-weeks, or as late effects, in which case disease onset can take years. In this work, we integrate clinical risk factors, radiotherapy treatment plan parameters, and patient specific biological variables together to construct models that more accurately predict such side effects. We make use of a well-established data-driven modelling approach based on logistic regression and optimised specifically for radiotherapy-related data-mining.

Teaching Lecture: Ensuring Quality in an Image Guidance Era

SP-0008 Ensuring Quality in an Image Guidance Era E. Miles 1 1 Mount Vernon Hospital, Academic Physics, Northwood Middlesex, United Kingdom Abstract text The radiotherapy process is a series of events during which discrepancies between the planned treatment and actual treatment delivered can occur. This necessitates a comprehensive quality assurance (QA) programme, including regular quality control (QC) checks and audits. As more advanced technology is introduced in the clinical setting, QA activities must continually evolve to provide a safe framework for implementation of technical radiotherapy. With image guided and adaptive strategies being increasingly employed to ensure accurate delivery of treatment in scenarios such as dose escalation and hypofractionation; techniques must be implemented in a safe and effective manner. QA in the clinical trial arena has played a leading role in striving for accuracy and consistency of radiotherapy treatment delivery through monitoring protocol compliance in a multi-centre setting. Clinical trials can also evaluate the feasibility and effectiveness of a new technology. A comprehensive trial QA programme not only accredits centres for recruitment to a trial but also benefits the general standard of radiotherapy delivered. This presentation will aim to demonstrate how we can extend the clinical trial QA experience to routine practice to ensure quality of image guidance through discussing examples of clinical trial benchmarking and credentialing processes and their perceived impacts on clinical practice. SP-0009 Clinical applications of AI for Radiation Oncology J. Bibault 1 1 Université Paris Descartes - Hôpital Européen Georges Pompidou, Paris, France Abstract text The application areas of Artificial Intelligence (AI) in radiation oncology include image segmentation and detection, image phenotyping and radiomic signature discovery, clinical outcome prediction, and treatment planning automation. In this lecture, we will first explain the methods used in AI, such as K-Nearest Neighbor, Decision Trees, Support Vector Machines, and Artifical Neural Networks, with a focus on Deep Learning (DL). In the second part of the lecture, we will describe studies using DL for image segmentation, outcome prediction (toxicity, treatment response and survival) or treatment planning. Finally we will explain the limits of these methods and evaluate how they will tranform radiation oncology. SP-0010 Acceptance, commissioning, introduction, regulatory aspects and QA of AI W. Verbakel 1 1 VU University Medical Center, Radiation Oncology Department, Amsterdam, The Netherlands Symposium: Artificial intelligence in Radiation Oncology