ESTRO 2020 Abstract Book

S853 ESTRO 2020

MHD with PhTh was 4.9 Gy (2.9 Gy – 6.6 Gy) and with PrTh 0.88 GyRBE (0.11 GyRBE – 2.63 GyRBE). Conclusion Approximately 3.2% of the BC patients qualified for proton therapy, when the selection was made using an NTCP model for cardiac toxicity with a delta NTCP of ≥ 2%. PO-1574 Robustness of CT-based prostate radiomics features against artefacts from gold fiducial markers S. Osman 1 , S. Jain 1 , A.R. Hounsell 2 , K.M. Prise 1 , C.K. McGarry 2 1 Queen's University Belfast, Centre for Cancer Research & Cell Biology, Belfast, United Kingdom ; 2 Belfast Health and Social Care Trust, Radiotherapy Physics, Belfast, United Kingdom Purpose or Objective A previously published model demonstrated the value of CT-based radiomics for prostate cancer (PCa) risk stratification [1]. External validation studies are required. As many hospitals routinely employ (gold) fiducial markers (FM) based IGRT, we investigated the robustness of CT- based radiomics features in the presence of FM and the associated artefacts. Moreover, to determine the utility of robust features, they were employed in building a risk- group classification model (low vs high) applying the same steps in the original published model [1]. Material and Methods Forty CT scans of 20 PCa patients acquired before and after FM insertion were used in this study. For each patient, a prostate gland only (P) structure was contoured on pre-FM CT and on post-FM CT. Population averaged mean prostate Hounsfield Unit (P HU ± SD ) was calculated from pre-FM scans. To reduce the influence of artefacts, four more structures were created on each post-FM scan; excluding only the FM (P-FM); and extracting associated artefacts by restricting the HU inside the prostate to (P HU ±1SD), (P HU ±2SD) and (P HU ±3SD) via thresholding. Radiomics features based on intensity histograms and texture matrices (GLCM, GLRLM, GLSZM, GLDZM, NGTDM and NGLDM) were extracted from all contoured structure after image pre-processing [1]. Using the pre-FM scan as a reference, intra-class correlation coefficients (ICC) were computed to provide an estimate of robustness (ICC = 0 non-robust, ICC= 1 perfectly robust features). In this analysis, a threshold of ICC > 0.8 was considered to identify robust features. The testing cohort consist of 32 low-risk and 233 high-risk patients. Results On pre-FM scans, the P HU was 36.1±3.3HU and the largest variability (SD) within a patient’s prostate was 21.6HU. This value was subsequently used for thresholding (SD=21.6). The structure that produced the largest number of robust features (306/2808) was (P HU ±1SD). Features extracted from (P) and (P-FM) produced the least number of robust features, Figure (1). (1)Darby et al. NEJM 2013;368(11):987-998 (2)Korevaar et al. Radioth Oncol 2019, in press

In a previous test-retest analysis, 522 features were reported to be robust [1]. The overlap between these 522 features and the currently identified features (306) was 133 features. Using only these features, we trained classifiers for risk group. The results are presented in figure 2. As for the original published model [1], the best classifiers (tested on unseen test dataset) were obtained using classifiers built using data augmentation to balance training data sets.