ESTRO 2021 Abstract Book

S1520

ESTRO 2021

of Oncology, IRCCS, Radiation Oncology , Milan, Italy; 6 IEO European Institute of Oncology, IRCCS, Molecular and Pharmaco-Epidemiology unit, Department of Experimental Oncology, Milan, Italy; 7 IEO, European Institute of Oncology IRCCS, Precision Imaging and Research Unit, Milan, Italy; 8 IEO European Institute of Oncology, IRCCS, Precision Imaging and Research Unit, Milan, Italy; 9 IEO European Institute of Oncology, IRCCS, Breast Imaging Division, Milan, Italy; 10 IEO European Institute of Oncology, IRCCS, Department of Radiology, Milan, Italy; 11 National Research Council, G. Monasterio Foundation, Pisa, Italy; 12 IEO European Institute of Oncology, IRCCSIEO European Institute of Oncology, IRCCS, Scientific Directorate, Milan, Italy Purpose or Objective Increasing efforts have been made to implement radiomics in the clinical practice. Features variability still represents one of the most critical pitfalls in the radiomic workflow, especially in magnetic resonance imaging (MRI). The aim of this study is to quantify the features robustness when using different radiomic packages and to provide harmonized parameters settings to allow for the greatest possible consistency among them. Materials and Methods Data from fluid-attenuated inversion recovery (FLAIR) MRI images of two patients diagnosed with glioblastoma who underwent surgery and chemoradiotherapy were included in the analysis. A subset of 20 radiomic features was selected according to the literature, and then extracted using Imaging Biomarker EXplorer (IBEX v1.0 beta) and PyRadiomics v3.0.1, with their default parameters settings. Harmonization methodology was derived from the study by Joseph James Foy (doi:10.6082/uchicago.2253). The harmonization effect was rated both intra-platforms and inter-platforms, through a Harmonization Necessity Factor (HNF) and PyRadiomics-Related Deviation (PyRD), respectively. HNF values of 2 or higher were associated with high sensitivity to harmonization, while PyRD values ≤0.01 were associated with excellent agreement in feature values among platforms.

Results Harmonized parameters settings led to the greatest coherence, in terms of lowered PyRD values compared to default ones (Fig. 1). For each patient 16/20 features reflected excellent PyRD agreement in feature values among packages, with only sphericity , Long Run Emphasis (LRE) , Long Run Low Gray Level Emphasis (LRLGLE) and coarseness having a PyRD value in the order of magnitude of 0.1. Based on HNF values, harmonization process had no effect on shape and first order features while C lusterProminence and Contrast showed to be the most sensitive, both in PyRadiomics and IBEX. Conversely, harmonization process had little impact on Short Run Emphasis (SRE) , Inverse Difference Moment Normalized (IDMN) and Correlation .

Conclusion With default settings, most features showed significant differences in computed values between packages. First order features proved excellent agreement based on PyRD factor, while second-order features showed a relatively poor agreement between packages. Overall, harmonization allowed to achieve a higher degree of consistency. Our preliminary findings suggest that features harmonization could be beneficial in the radiomic workflow and foster results comparison across different centers towards clinical application. PO-1795 diffusion model based target definition for high-grade gliomas: robustness analysis W. Häger 1 , M. Lazzeroni 1 , M. Astaraki 2 , I. Toma-Dasu 3 1 Stockholm University, Department of Physics, Stockholm, Sweden; 2 Royal Institute of Technology, Department of Biomedical Engineering and Health Systems, Huddinge, Sweden; 3 Karolinska Institute, Department of Oncology and Pathology, Stockholm, Sweden