ESTRO 2020 Abstract book

S303 ESTRO 2020

networks (GAN) have been successfully applied to CBCT correction. Researchers have trained CNNs using both projection data and reconstructed images, often relying on DIR and MC to generate ground truth data. By relying on data augmentation strategies, CNNs have been trained with datasets typically featuring <100 patients, yielding convincing results for both photon and proton therapy dose calculation, but also for segmentation of CBCT images. This contribution will aim at presenting a comprehensive overview of this relatively young field. 1. Thing RS, et al. Phys Med Biol. 2016;61(15):5781- 802. 2. Peroni M, et al. Int J Radiat Oncol Biol Phys. 2012;84(3):e427-33. 3. Park YK, et al. Med Phys. 2015;42(8):4449-59. SP-0512 On the crucial importance of image preprocessing and harmonization for robust AI-based models I. Buvat 1 1 Université Paris-Sud- Université Paris-Saclay, LITO- UMR1023 Inserm- Institut Curie, Paris, France Abstract text Artificial intelligence-based models offer great potential to make the most of medical images for patient management and treatment planning. These models are based on engineered or deep features calculated from medical images, aka radiomic features, and designed in such a way that they capture the most relevant information present in the images to perform a task (for instance delineate a tumor or organ at risk), to make a prediction or to classify patients. AI-based radiomic models are derived through a training process most often involving annotated data, ie images associated with a (surrogate) ground truth. They should then be validated on data that have not been used to create the models, and for which a ground truth is also available to characterize the performance of the model. One of the main challenges faced by radiomics lies in the variability in image properties (eg, spatial resolution, signal-to-noise ratio) and quality, as a function of the specific acquisition device and protocol used to produce the images. Such variability has been shown to penalize the ability of the model to generalize, that is to perform well on data acquired in different conditions (Reuzé et al, Oncotarget 2017). To tackle that problem, several approaches can be used, among which training the model based on a large variety of data encompassing a broad range of acquisition conditions, harmonizing the images before training, harmonizing the features before designing the model, or fine tuning the model to the specificities of each dataset. The relevance and applicability of these different approaches will be compared and discussed, and practical solutions will be presented with examples in CT, PET and MR imaging. The impact of harmonization procedures on the robustness of AI-based radiomic models will be described and illustrated.

Symposium: Individualised radiotherapy

SP-0513 PET-based dose painting in head and neck cancer M.E. Evensen 1 , E. Dale 1 1 Oslo University Hospital, Department Of Oncology, Oslo, Norway Abstract text Traditionally all parts of a tumor have been treated with the same radiation dose, but the advent of intensity modulated radiation therapy has made it possible to modulate the dose within the tumor. Dose painting uses biomarkers to identify areas of the tumor believed to be aggressive or treatment resistant and increase the radiation dose to these areas of the tumor. PET is the most commonly used biomarker for dose painting today. Different strategies have been suggested as ways to escalate the dose within the tumor; dose painting by numbers, where each voxel is assigned a dose related to the PET intensity of the voxel, and dose painting by contours, where all voxels above a threshold value are escalated. A third option is a hybrid approach, using dose painting by contours with several threshold values. The end result is a personalized intra-tumor dose distribution. Head and neck cancer patients with squamous cell carcinoma (HNSCC) not related to human papilloma virus (HPV) have a significantly worse outcome than patients with HPV-related disease. Failures and recurrences are also usually located within the original gross tumor volume, and with even higher likelihood within the high uptake region defined by fluorodeoxyglucose (FDG) positron emission tomography (PET). It is hypothesized that high standard uptake values (SUV) represent aggressive sub-sets of the tumor. These areas are the primary target of dose painting in head and neck cancer. Since the mid-2000s there has been published data from a few studies on dose painting in head and neck cancer, primarily using FDG-PET and dose painting by numbers. Results have been promising, but late side-effects, in particular late grade 3+ mucosal ulcers, have been seen in some patients. In addition, current treatment planning systems do not directly support dose painting by numbers. The RIDPAINT (NCT02921581) and RADPAINT (NTC03847480) studies are two phase I feasibility studies, for primary and recurrent HNSCC, where a hybrid approach is used. The RADPAINT study also aims to reduce the dose to the mucosa in accordance with recommendations from previous studies. This will hopefully reduce the incidence of late side-effects. A dosimetric study was done along with the first inclusions into the RADPAINT study, to identify possible increases in dose to organs at risk, with particular focus on the swallowing structures. Results from this study, along with early results from the RADPAINT study will be presented, in addition to results from the RIDPAINT study.