OA071] O-RAW: Ontology-guided radiomics analysis workflow

Abstract

Purpose Radiomics is high-throughput computerized tumour feature extraction from medical images. This has shown potential for quantifying tumour phenotype and predicting treatment response. Three major challenges impede the pace of radiomics research and clinical adoption: (i) lack of standardized methodology for radiomics analyses, (ii) lack of universal lexicon to denote features that are semantically equivalent and (iii) flat tables for radiomics output do not sufficiently capture the methodological steps that affect feature values. These barriers hamper multi-centre validation studies applying subtly different imaging protocols, pre-processing steps and extraction software. We propose an open-source Ontology-guided Radiomics Analysis Workflow (O-RAW) to address the above challenges. Methods O-RAW was developed in Python, which comprises three phases and uses two open-source component libraries (Py-rex and Pyradiomics). First, Py-rex takes standard DICOM-RT inputs (DICOM images and an RTSTRUCT file) and parses them as numpy arrays of voxel intensities and a binary mask for the volume of interest (VOI). Next, the numpy arrays are passed to Pyradiomics performing the feature extraction and returns a dictionary object to Py-rex. Lastly, Py-rex parses the dictionary as a W3C–compliant semantic web “triple store” (i.e., list of subject-predicate-object statements) with relevant semantic metalabels drawn from the Radiation Oncology Ontology and Radiomics Ontology. The output is published on a http-accessible endpoint, and can be examined via SPARQL queries. In this study, we demonstrate how DICOM-RT data can be downloaded directly from a public image repository and analysed for correlation. Results We showed that O-RAW executed efficiently on two lung cancer datasets including 21 and 32 lung cancer patients, with a median 12 and 2 VOIs per patient respectively, on a laptop running Windows 7 operating system and 8 GB RAM. The average execution time of feature extraction was 5.3 ± 0.7 s per VOI. Conclusions We successfully implemented O-RAW for radiomics analysis from radiotherapy-based images to semantic triples. With the guide of ontology, O-RAW can create findable, accessible, interoperable and reusable (FAIR) data. Its practicability and?exibility can greatly increase the development of radiomics research and finally transfer to clinical practice. We validated O-RAW on a number of publicly available image collections.