ESTRO 2020 Abstract book

S885 ESTRO 2020

accessible, interoperable and re-usable (FAIR) through a public semantic web access point (http://sparql.cancerdata.org) Material and Methods Overall survival intervals (days since start of radiotherapy) have been updated through the Dutch national registry under an internal ethics board-approved request. Spatially incorrect offsets of the Primary Gross Tumour Volume (“GTV-1”) regions of interest (ROIs) in the Lung1 set were amended in The Cancer Imaging Archive (TCIA) collection . Image features from each GTV-1 were extracted using an ontology-guided radiomics workflow (O-RAW; link) and published in Resource Descriptor Format (RDF) consistent with the Image Biomarker Standardization Initiaitive (IBSI) through an open Radiomics Ontology. DICOM metadata as RDF was extracted using a research version of Semantic DICOM (SoHard, GmbH, Fuerth; Germany). Clinical data was published in RDF using the Radiation Oncology Ontology. Example queries were tested, which verified that the SPARQL endpoint was accessible. The digital artefact has been registered on Zenodo with detailed descriptions. Results The SPARQL endpoint is publicly accessible. One of the major advantages of the semantic web is that all SPARQL queries are automatically federated, thus we can effortlessly cross-reference into the clinical, dicom and radiomics data. Purely as a hypothetical example, we crafted a query which tabulated the radiomics feature for every patient along with the scan milliampere-seconds (mAs) settings. We then analysed the query result to generate a plot of Intra-class Correlation Coefficient (ICC) distribution as a function of feature class and mAs range (see Figure 1).

Netherlands, Department of Radiation Oncology MAASTRO, Maastricht, The Netherlands

Purpose or Objective In this study, we compared clinical variables and quantitative imaging features (radiomic and deep learning features) fitted into a pooling-based feature selection method for 2 year survival prediction of lung cancer patients. Material and Methods 421 non-small cell lung cancer (NSCLC) patients treated with radical radiotherapy and chemo-radiation were collected in this study. Radiotherapy treatment planning DICOM CT images were used. For clinical variables, clinical TNM stages were included. For quantitative imaging features, a total of 1092 handcrafted radiomic features consisting of histogram statistics, shape, texture and features by Wavelet and Laplacian of Gaussian filtering, were extracted from the GTV via an open-source radiomics package O-RAW. Second, 256 deep learning radiomic features were extracted from the 3D isotropic patches of dimension 50 × 50 × 50 around each tumour center of mass. Both handcrafted radiomics and deep learning radiomics were passed into a pooling feature selection workflow. Finally, seven prediction models were developed using (1) TNM stage, (2) tumour volume alone, (3) selected radiomic features, (4) radiomic features combined with TNM stage, (5) radiomic features combined with tumour volume, (6) deep learning radiomics + pooling feature selection, and (7) deep learning radiomics + fully-connected convolutional neural networks (CNN) model respectively. The area under curve (AUC) along with standard deviation (sd) in a 5-folder cross validation (CV) was determined to assess the model discrimination. Results After implementing the pooling feature selection method, 4 radiomic features were remained composing the finial signature. TNM stage, volume and the signature achieved AUCs of 0.516 ± 0.065, 0.642 ± 0.092 and 0.670 ± 0.044 respectively. When adding TNM stage and volume to the radiomic signature model, mix models reached AUCs of 0.657 ± 0.039 and 0.660 ±0.065 respectively. The deep learning radiomic features model, consisting of 15 deep features, reached an AUC of 0.641 ± 0.057 for 5-fold CV. When fitting the 256 dimensional deep features into a fully-connected CNN classifier, the model reached a AUC of 0.621 ± 0.029. The whisker boxplot of AUC performance for 7 models were included in Figure 1.

. Conclusion

We successfully generated separate RDF repositories of clinical, DICOM and radiomics data and published these on an open access SPARQL endpoint. Queries can be generated which simultaneously looks in all three repositories, thus taking advantage of the semantic linking between the data elements. We expect that having this data resource publicly available will promote investigation of radiomics features repeatability and reproducibility. References 1: PMID: 24892406 2: PMID: 29092951 Keywords: Radiomics, public datasets, reproducibility, FAIR data PO-1544 Comparing Clinical Variables and Quantitative Imaging Features for Lung Cancer Survival Prediction C. Zhang 1 , Z. Shi 1 , C. Zhu 1 , P. Kalendralis 1 , I. Bermejo 1 , L. Wee 1 , A. Dekker 1 1 GROW School for Oncology and Developmental Biology- Maastricht University Medical Centre+- Maastricht- The

Conclusion In this study, we compared clinical variables and quantitative imaging features (radiomic and deep learning