ESTRO 2021 Abstract Book

S526

ESTRO 2021

size of just one slice. A ROC analysis yielded an overall AUC of 0.985 with a detection accuracy of 97.6%. Moreover, the subsequent inpainting step improves the image quality of artifact-affected slices considerably, leading to a 41.5% decreased RMSE and 78.9% increased NCC averaged over all test patients. In particular, the inpainting network managed to restore tumors directly affected by interpolation artifacts (c.f. Figure 1), since the process is conditioned on the average image. Building upon these promising results, the presented method should be extended to double structure artifacts in future research.

Conclusion Deep learning-based artefact detection and conditional inpainting is a promising approach to replace non- present anatomical information in linear interpolation artifact areas, exploiting available patient-specific image information.

Poster highlights: Poster Highlights 25: Prostate

PH-0656 Prediction of toxicity after prostate cancer RT: the value of a SNP-interaction polygenic risk score T. Rancati 1 , M. Massi 2 , N. Franco 2 , B. Avuzzi 3 , D. Azria 4 , A. Choudhury 5 , A. Cicchetti 1 , D.R. Dirk 6 , A. Dunning 7 , R. Elliott 5 , F. Ieva 2 , S. Kerns 8 , M. Lambrecht 9 , A. Manzoni 2 , A. Paganoni 2 , B. Rosenstein 10 , P. Seibold 11 , E. Sperk 12 , C. Talbot 13 , A. Vega 14 , L. Veldeman 15 , P. Zunino 2 , A. Webb 13 , J. Chang-Claude 11 , C. West 5 1 Fondazione IRCCS Istituto Nazionale dei Tumori, Prostate Cancer Program, Milan, Italy; 2 Politecnico di Milano, MOX Laboratory, Math Department, Milan, Italy; 3 Fondazione IRCCS Istituto Nazionale dei Tumori, Radiation Oncology 1, Milan, Italy; 4 Montpellier Cancer Institute, Univ Montpellier MUSE, Grant INCa_Inserm_DGOS_12553, Inserm U1194, Radiation Oncology, Montpellier, France; 5 University of Manchester, Manchester Academic Health Science Centre, Christie Hospital, Translational Radiobiology Group, Division of Cancer Sciences, Manchester, United Kingdom; 6 GROW Institute for Oncology and Developmental Biology,, Radiation Oncology (Maastro), Maastricht, The Netherlands; 7 Centre for Cancer Genetic Epidemiology, University of Cambridge, Strangeways Research Labs, Oncology, Cambridge, United Kingdom; 8 University of Rochester Medical Center, Radiation Oncology and Surgery, Rochester, USA; 9 University Hospitals Leuven, Radiation Oncology, Leuven, Belgium; 10 Icahn School of Medicine at Mount Sinai, Genetics and Genomic Sciences and Radiation Oncology, New York, USA; 11 German Cancer Research Center (DKFZ), Cancer Epidemiology, Heidelberg, Germany; 12 Universitätsmedizin Mannheim, Medical Faculty Mannheim, Radiation Oncology, Mannheim, , Germany; 13 Leicester Cancer Research Centre, University of Leicester, Genetics and Genome Biology, Leicester, United Kingdom; 14 Fundación Pública Galega de Medicina Xenómica, Grupo de Medicina Xenómica (USC), Santiago de Compostela, Spain; 15 Ghent University Hospital, Radiation Oncology, Ghent, Belgium Purpose or Objective An international, prospective cohort study recruited prostate cancer patients (pts) in 8 countries (April2014- March2017). It was aimed at multinational validation of clinical/dosimetric/genetic risk factors that predict late toxicity following radiotherapy (RT). The purpose here was to propose a novel scoring method to summarize genetic information that incorporates epistatic effects (i.e. a polygenic risk score, PRS, that for the first time incorporates SNP-SNP interactions, PRSi) and to investigate the benefit of adding PRSi in risk prediction models. Materials and Methods 1808 pts were enrolled. RT was prescribed according to local regimens, but centres used standardised data collection. Grade≥1 (G1+) rectal bleeding (G0 at baseline), G2+ urinary frequency (GO/G1 at baseline), G1+ weak stream (G0 at baseline), G1+ haematuria (G0 at baseline), G2+ nocturia (G0/G1 at baseline) were considered as separate endpoints. Clinical/dosimetric models for these late endpoints including validated features (already published and confirmed in this cohort) were presented at ESTRO & ASTRO 2019 (summary in Fig 1). A pool of 43 SNPs associated with late RT toxicity from the literature was tested, and a deep sparse autoencoder method identified the SNPs affecting the toxicity risk (n=13; Fig 2a). A new method for