ESTRO 2020 Abstract Book

S827 ESTRO 2020

Conclusion Conclusion: Radiomic features from dose maps seem to add a significant value to usual NTCP models and could help better assess the risk of clinically relevant radiation induced lung toxicity. Evaluation of the trained models in another cohort is currently undergoing. PO-1531 Publishing linked and FAIR radiomics data in radiation oncology via ontologies and Semantic Web A. Traverso 1 , M. Vallieres 2 , J. Van Soest 1 , L. Wee 1 , O. Morin 3 , A. Dekker 1 1 Maastricht Radiation Oncology MAASTRO clinic, Radiotherapy, Maastricht, The Netherlands ; 2 McGill University, Medical Physics, Montreal, Canada ; 3 University California San Francisco, Radiation Oncology, San Francisco, USA Purpose or Objective Radiomics showed promising results in radiation oncology. However, most of the radiomics studies remain difficult to reproduce because of poor quality of reporting and the adoption of the so-called black-box approach. This lack of transparency is limiting the generalizability and validity of radiomics-based models. In this study, we present a proof- of-concept study using our newly developed radiomics ontology (RO), combined with Semantic Web technologies, as instrument for enabling interoperability of radiomics data following FAIR (Findable Accessible Interoperable We developed the radiomics ontology (RO, available on BioPortal): 458 classes and 76 predicates covering the whole spectrum of the workflow of radiomics computation, as presented in the IBSI (Image Biomarker Standardization Initiative). As a proof of concept, two institutions used two different open source radiomics packages to extract features from gross tumor volumes scans from publicly available CT scans of lung cancer patients. Features and metadata were converted from relational databases to RDF (Resource Description Framework) triples and uploaded to a SPARQL endpoint (Figure1a), as per FAIR guidelines. Reusable) principles. Material and Methods

Conclusion By embedding FAIR principles in radiomics computations, transparency of radiomics studies is augmented. The Radiomics Ontology represents a natural extension of the IBSI effort, by providing a universal language for reporting radiomics data and metadata. Finally, this ontology represents the backbone for including radiomics to distributed learning networks. PO-1532 Prediction of Lymph Node Metastases via PET Radiomics of Primary Tumour in Esophageal Adenocarcinoma Z. Shi 1 , C. Zhang 1 , P. Kalendralis 1 , P. Whybra 2 , C. Parkinson 2 , M. Berbee 1 , E. Spezi 2 , A. Roberts 3 , A. Christian 4 , T. Crosby 5 , A. Dekker 1 , L. Wee 1 , K.G. Foley 6 1 GROW School for Oncology and Developmental Biology- Maastricht University Medical Centre+, Department of Radiation Oncology MAASTRO CLINIC, Maastricht, The Netherlands ; 2 Cardiff University, School of Engineering, Cardiff, United Kingdom ; 3 University Hospital of Wales, Department of Radiology, Cardiff, United Kingdom ; 4 University Hospital of Wales, Department of Pathology, Cardiff, United Kingdom ; 5 Velindre Cancer Centre, department of clinical oncology, Cadiff, United Kingdom ; 6 Velindre Cancer Centre, department of clinical radiology, Cadiff, United Kingdom Purpose or Objective To improve clinical lymph node staging (cN-stage) in esophageal adenocarcinoma by developing and externally validating three prediction models with 1) clinical variables 2) positron emission tomography (PET) radiomics and 3) a combined clinical and radiomics model. Material and Methods Consecutive patients with fluorodeoxyglucose (FDG) avid tumours treated with neo-adjuvant therapy between 2010 and 2016 in two international centres (n=130 and n=60, respectively) were included. Four clinical variables (age, gender, clinical T-stage and tumour regression grade) and PET radiomics derived from the primary tumour were used to construct the three models. Accuracy, sensitivity, specificity, positive predictive value, negative predictive value, area under curve (AUC), discrimination and calibration were calculated for each model. The prognostic significance was also assessed.

Results Each of the users could independently query all the features generated from the two Software, without having any prior knowledge of the original labels used to store features and associated computational details. Using SPARQL queries, we could reveal the parameters accounting for major differences in feature values as a function of mapped metadata, enabling interoperability between different radiomics computational packages. Finally, radiomics features were linked to corresponding clinical data as shown in Figure 1b. The group has released a python package of the whole proof of concept, which allows any radiomics user to adopt FAIR principles for its study.