ESTRO 2020 Abstract Book

S847 ESTRO 2020

Results

robustness was classified using pre-determined thresholds. Results Mean DSC was 0.65±0.12 (range: 0.19-0.77). 49% of the participant group achieved a DSC >0.7 and the auto- segmentation DSC was 0.75. 45% and 34% of features were deemed to be robust with a mean paired percentage difference <5%, for the auto-segmentation and manual delineations respectively. The majority of these were texture features from the gray-run length matrix family. 5% of features from the auto-segmentation and 10% of features from the manual contours were deemed to be unstable with a mean paired percentage difference >50%. Conclusion Delineation of the parotid gland is challenging and inter- observer delineation variability is apparent. This study was able to isolate a number of radiomics features which are more robust to these uncertainties. These specific features now need to be investigated to ascertain their clinical value in prediction of xerostomia, with a view to adapt treatment based on quantitative imaging analysis. PO-1565 Multiparametric MRI radiomics model to predict overall survival in Glioblastoma Multiforme E. Kolozsi 1 , J. Powell 2 , C. Piazzese 1 , S. Thomas 2 , J. Staffurth 2 , E. Spezi 1 1 Cardiff University- School of Engineering, Biomedical Engineering, Cardiff, United Kingdom ; 2 Velindre Cancer Centre, Department of Clinical Oncology, Cardiff, United Kingdom Purpose or Objective High-grade glioma (HGG) is the commonest primary brain tumour in adults, for which current treatment options are limited with a poor overall prognosis. Risks associated with brain biopsy depend on tumour location and radiomics may allow non-invasive tumour assessment to avoid complications. Radiomics refers to the extraction and analysis of quantitative imaging features from medical images and offers an innovative approach to address diagnostic and prognostic challenges in HGG. The present study aims to build a radiomics model from magnetic resonance imaging (MRI) to improve overall survival prediction in Glioblastoma Multiforme (GBM). Material and Methods A cohort of 32 patients with pre-treatment T1-weighted, T1 post-contrast, T2-weighted and FLAIR MRI were recruited, clinical variables included gender and age. Prior to feature extraction image intensities were normalized between µ±3σ 1 and then quantized to 32, between 1 and 2 32 . A total of 182 standardised features 2 were extracted from the gross tumour volume (GTV) using an in-house software, Spaarc Pipeline for Automated Analysis and Radiomic Computing (SPAARC) 3 . Intra-class correlation coefficient (ICC) was used to detect stable features among MRI sequences. In order to build a multivariate model which predicts overall survival in GBM patients, feature selection was required to strengthen reliability. The Boruta machine learning-based algorithm 4 was used to capture statistically significant features with respect to survival outcome variable.

Figure 1: One axial slice of the frequency and 2 yrs OS and 5 yrs OS frequency weighted cumulative status (fwCS) map of the SAKK cohort. An increased change from controlled to non-controlled voxels can be observed at the main right bronchus when comparing at 5 yrs to 2 yrs OS. PT center of mass location from the in-house software agreed within +/- 3 mm Euclidean distance with the MIM software. The largest mean increase of fwCS value at 2 yrs and 5 yrs OS was observed at the heart and bronchi region for the SAKK cohort (Figure 1). The overlap of the maps for both cohorts is high (only 10% of PTs of the internal cohort were not covered by the SAKK cohort). At 2 yrs OS, the mean fwCS values indicate highest risk at the left-posterior chest wall for both cohorts and lowest risk for left anterior chest wall for the SAKK and heart region for the internal cohort. Univariate analysis to predict 2 yrs OS showed good performance, the best performing feature was robust mean absolute deviation (AUC = 0.75, p = 0.078). Conclusion This explanatory analysis quantifies the value of primary tumor location for OS prediction. Additional data is needed to provide more conclusive results. PO-1564 Influence of inter-observer delineation variability on radiomics features of the parotid gland E. Forde 1 , M. Leech 1 , C. Robert 2 , E. Herron 3 , L. Marignol 1 1 Trinity College Dublin, Discipline of Radiation Therapy, Dublin, Ireland ; 2 Gustave Roussy- Universite Paris- Saclay, Department of Medical Physics, Villejuif, France ; 3 Trinity College Dublin, Department of Psychiatry, Dublin, Ireland Purpose or Objective Previous research has demonstrated baseline radiomics features of the parotid glands are informative for predicting xerostomia. Whilst these studies are useful they fail to consider inherent inter-observer delineation variability. Other studies have investigated the impact of such variability on radiomics features; however, analysis has been limited to target delineation and variability in organ at risk contouring has not yet been investigated. Furthermore, existing literature is limited by the relatively small number of expert observers; possibly not reflecting practice in a true clinical environment. This study aims to identify the more robust features, defining the right parotid gland as the region of interest, taking data from a large number of observers. Material and Methods This research was a secondary analysis of anonymous data obtained from an ESTRO online delineation workshop. Participants (n=40) were provided with a common CT data set and, using FALCON software, were asked to delineate the organs at risk within the head and neck. Inter-observer variability for the right parotid gland was quantified with the DICE similarity coefficient (DSC); where the contour created by the workshop’s instructor was used as the reference (expert) contour. For comparison, an additional contour was generated using Varian SmartSegmentation (v15.5). Radiomics features were extracted using LIFEx software, including four shape features, six histogram features, and 32 texture features. The absolute mean paired percentage difference in feature values from the expert and participants were calculated. Feature