ESTRO 2020 Abstract book

S302 ESTRO 2020

tradeoff between expanding the clinical target volume (CTV) for better coverage of potential tumor spread in surrounding healthy tissues, and sparing of OARs. In the first part of this talk we will present rapid methods to automatically expand the CTV beyond the gross tumor volume (GTV), while respecting automatically defined (with deep learning) anatomic barriers to tumor spread. The computational method is based on the Dijkstra shortest path algorithm. In the second part we will translate calculated distances from the GTV into probabilities of tumor spread, based on limited data from biopsies and clinical experience. Instead of defining a binary CTV, we obtain a continuous Clinical Target Distribution (CTD). How to integrate this probability distribution in treatment planning optimization algorithms is discussed in the third part of the presentation. The simplest approach uses a probability-weighted sum of voxel-wise objectives. The results demonstrate “smarter” sparing of OARs based on their proximity to areas of (uncertain) tumor infiltration. Clinical example cases include glioma and sarcoma. In conclusion, we demonstrate how to generalize the concept of MCO to CTV definition. SP-0509 Integrating uncertainties in target definitions and dose prescription R. Jeraj 1 , P. Ferjancic 2 1 university Of Ljubljana, Faculty Of Mathematics And Physics, Ljubljana, Slovenia ; 2 university Of Wisconsin, Medical Physics, Madison, Usa Abstract text There are many uncertainties that affect target definition and need to be accounted for. Broadly they fall into two large categories: (1) uncertainties of tumor presence, which originate from imaging uncertainties and uncertainties related to microscopic infiltrations, and (2) uncertainties for "hitting" the target, which include residual position uncertainties and motion uncertainties. In this talk we will present and discuss: (1) Sources of uncertainties that affect target definition, (2) Impact of uncertainties on target definition and (3) Clinical examples of integrating uncertainties in target definition. SP-0510 Accelerating MRI reconstruction with deep learning K. Lønning 1 , M.W.A. Caan 2 , J. Sonke 1 1 netherlands Cancer Institute, Radiation Oncology, Amsterdam, The Netherlands ; 2 amsterdam Umc, Biomedical Engineering And Physics, Amsterdam, The Netherlands Abstract text Due to the high contrast in soft tissue inherent to the modality, Magnetic Resonance Imaging (MRI) is well suited for locating and tracking tumors, as well as distinguishing organs at risk. A problem with MRI, however, is that scans can be time consuming. For radiotherapy in particular, where treatment stands to benefit from online image guidance, a reduction of scan times is necessary to exploit the true potential of MRI. A common strategy for accelerating MRI is to sparsely sample the data necessary to create an image, after which an algorithm is applied to reconstruct the sub-sampling artifacts. Compressed Joint Symposium: ESTRO-EFOMP: Artificial intelligence and image quality (with focus on benefits for RT)

Sensing (CS) is one such algorithm widely used in radiology for reconstructing sparsely sampled MRI scans, but its speed of inference can be prohibitively slow if images are needed before the patient leaves the scan table, as the case is in radiotherapy applications. In recent years, the use of data-driven deep neural networks have led to not only a reduction in reconstruction times, but also an improvement in image quality compared to CS. Several such networks have been developed thus far, most of which deviate from the U-net architecture typically used for tasks where both the network’s input and output are images. Instead, the iterative scheme behind CS is molded into a neural network, as this better exploits the underlying physics of the accelerated MRI measurement process. The result is robust networks that have been shown to faithfully reconstruct new data after training on a very limited number of scans. Their application to the clinic promises to improve overall efficiency, and, more importantly, they should allow for motion monitoring at an increased rate. That being said, most models are currently confined to operate on regions with little to no motion during the scan itself, limiting their use in radiotherapy, where estimating a tumor’s position through respiratory cycles, cardiac cycles, or other sources of short-term motion is often necessary. Future research can therefore be expected to explore deep learning models designed to tackle dynamic scan protocols, with the potential for further imaging acceleration by exploiting temporal correlations. This presentation will give a summary of deep learning methods currently used in research, followed by an attempt to predict the sort of applications we may see evolve from them in the near future, including an overview of the challenges the field is faced with in order to get there. SP-0511 Improving image quality of CBCT using AI G. Landry 1 , C. Kurz 1 1 university Hospital- Lmu Munich, Department Of Radiation Oncology, Munich, Germany Abstract text It is well understood that while extremely valuable for bony anatomy registration, gantry-mounted cone beam computed tomography (CBCT) image quality is lower than diagnostic computed tomography (CT), mainly due to the detection of patient scatter. The lower image quality makes uncorrected CBCT images typically unsuitable for tumor visualization, delineation and dose calculation. Several methods have been developed to improve image quality, with the goal of realizing similar dose calculation accuracy as achievable on CT, for both photon and proton therapy. Well published approaches include the use of Monte Carlo (MC) simulation to estimate the patient scatter [1], or the use of deformable image registration (DIR) to deform the CT to the CBCT to create a virtual CT [2], either as an end in itself or to generate a prior for projection correction algorithms [3]. Both MC and DIR are however time-consuming algorithms and are not easily made compatible with the requirements of online dosimetric evaluation of anatomical changes. By now several researchers have realized that artificial intelligence (AI) tools may in certain cases offer significant speed up in radiotherapy; mirroring broader general trends. While some studies report the use of classical AI tools such as random forests, recent contributions have focused on the use of deep-learning using convolutional neural networks (CNNs). Established CNN architectures such as U shaped CNNs (U-Net) and generative adversarial