ESTRO 2023 - Abstract Book

S12

Saturday 13 May

ESTRO 2023

SP-0033 Overview of the main families of available solutions S. Pallotta 1 1 University of Florence, Experimental and Clinical Biomedical Sciences "Mario Serio", Italy, Italy

Abstract Text The tipical radiotherapy (RT) pipeline consists of the following steps: image acquisition, target and organ-at-risk (OARs) segmentation, treatment plan generation, quality assurance (QA) and treatment delivery. Many of these processes are time-consuming and provide results that depend on the operators' skills and experience. By automating routine tasks, artificial intelligence (AI) techniques can help standardize working methods, reduce the risk of errors, and improve process efficiency. I will review the primary families of available solutions, highlighting possible challenges and future developments. Atlas-based segmentation software, which uses deformable registration to propagate the labelled structures in the atlas image onto the target image, has been successfully used for automatic OAR segmentation. Usually, they work well for some data sets or patients, but the results are not always generalizable. The application of AI techniques based on deep learning (DL) to the segmentation of medical images has recently enjoyed widespread success. DL approaches gained popularity showing promising results for OAR segmentation and, recently, for gross tumor volumes segmentation [1]. The primary type of imaging used for RT treatment planning is Computed Tomography (CT). It provides accurate information about the patient's geometry and allows direct conversion of the electron density required for dose calculation. Recently, MRI-only based RT has been proposed to simplify the workflow and reduce the overall treatment costs and workload. In addition, the implementation of an MRI-only workflow can be beneficial for MRI-guided treatment techniques such as the MRI-linear accelerator. MR can be converted into a CT-equivalent representation, creating a synthetic CT, according to three main approaches. These consist of bulk density assignments, atlas-based techniques, and voxel-based techniques [2]. DL-based methods to generate synthetic CTs have also received considerable attention as an alternative to classical methods [3]. Radiotherapy treatment can also be automated by following different strategies. The most popular automated treatment planning (ATP) approaches are three [4]. Those that use prior knowledge to predict the achievable dose in a new patient belong to the knowledge-based planning category. All those systems that automate the iterative process of optimizing objectives and constraints belong to the iterative protocol-based optimization category. Finally, in multi-criteria optimization, the optimal compromise between target coverage and sparing of all normal tissues is sought, according to the concept of 'optimal Pareto solution'. In addition, some DL-based ATP techniques have been proposed using various kinds of neural networks. The main implementations focused on automated beam orientation selection, automated dose map prediction, and automated fluence map and delivery parameters generation [5]. Treatment planning and delivery techniques are subject to a wide range of uncertainties that may impact patient-specific treatment verification results. The principal causes of disagreement between calculated and measured dose maps are due to the characteristics of the detector used, the beam modelling in the TPS, the dose calculation engine, the beam output and profile, and above all, the complexity of the plan. Depending on this, it may be complex for the linac to deliver the treatment as planned by the TPS. Assuming that less complex plans are at low risk of failure, predictive algorithms based on DL strategies could be used to identify complex plans that need to be verified [7]. Moreover, AI-based systems seem to be promising in enabling efficient triage for problem-solving and setting preventive actions also for linac QA. Finally, despite the efforts made to develop innovative machine learning and DL algorithms, the path to quality assurance is not yet fully mapped out. To ensure that a model based on AI techniques meets clinical expectations, appropriate measures are needed to intercept any conditions that could cause inaccurate results. This aspect leads towards a demand for AI systems that are not only of high quality but also interpretable [8]. [1] Savjani et al. Semin Radiat Oncol 2022 doi:10.1016/j.semradonc.2022.06.002. [2] Johnstone et al. IJROBP 2018. doi: 10.1016/j.ijrobp.2017.08.043. [3] Gurney-Champion et al. Semin Radiat Oncol 2022 doi:10.1016/j.semradonc.2022.06.007 [4] Hussein et al. BR J Radiol 2018. doi:10.1259/bjr.20180270. [5] Wang et al. Frontiers in Oncology 2020. Doi:10.3389/fonc.2020.580919

[6] Vandewinckele et al. Radiotherapy and Oncology 2020 doi:10.1016/j.radonc.2020.09.008 [7] Vandewinckele et al. Radiotherapy and Oncology 2020 doi.org/10.1016/j.radonc.2020.09.008 [8] Claessens et al. Semin Radiat Oncol 2022 doi.org/10.1016/j.semradonc.2022.06.011 SP-0034 Deep learning for segmentation and treatment planning for breast cancer patients C. Hurkmans 1 1 Catharina Hospital Eindhoven, Radiation Oncology, Eindhoven, The Netherlands

Abstract Text In this presentation the practical clinical implementation of deep learning (DL) based segmentation and planning for breast cancer treatment planning will be discussed. It will include all steps that one would need to take to use DL in clinical practice: pre-clinical training, evaluation and testing when building your own model or testing of a commercial model. Setting standards for comparisons with your own clinical data, both quantitatively and qualitatively, which includes e.g., DVH criteria, possible time gains, fraction of plans or segmentations that still need to be adjusted etc. Documentation for commissioning and the Medical Device Regulations, education and monitoring during clinical use. The changing roles of radiation oncologists, RTT and medical physicists will also be discussed.

SP-0035 From atlas-based knowledge to ML-based planning solutions C. Khamphan France

Abstract not available