ESTRO 37 Abstract book

S1186

ESTRO 37

correlation test was used to assess the correlation differences in radiomics features compared to the baseline value. Levene’s test was considered to assess any changes in the distribution of the differences with increase in the baseline value. Principal component analysis was also considered to determine the number of components required to describe 95% of feature space variation. Results Over half of considered features showed instability towards changes in voxel size, signal intensity bins and contours. More features were impacted by the considered changes in voxel size, than the considered changes in contours or signal intensity bins. The features demonstrating a significant correlation with Pearson’s test varied depending on the three imaging variation methods. Results from Levene’s test also varied from the correlation results and with the imaging variation which was being considered. Six principal components were required to describe 95% of the variation in the data. Conclusion The majority of radiomics feature values were impacted by changes in voxel size, signal intensity bin and small variations in contour, and might therefore may not be appropriate for cross institution decision models. Use of principal components could describe 95% of the feature variation, which is a significant reduction of feature space. Choice of radiomics features to be considered when developing outcome models must be undertaken cautiously. EP-2146 Feasibility of direct electron density determination using dual-energy cone beam CT L. Schröder 1 , U. Stanković 1 , M.F. Fast 1 , J.J. Sonke 1 1 Netherlands Cancer Institute, Radiation Oncology, Amsterdam, The Netherlands Purpose or Objective Establishing the relationship between the relative electron density (RED) and the Hounsfield units (HU) is an essential step in radiotherapy dose calculation. As the HU-RED relationship is not linear, it is often characterized using a piecewise linear model between the measured HU of ground-truth material with known RED. However, due to different effective atomic numbers (Z eff ), the same HU can represent different RED. Saito [1] has proposed a method for the diagnostic CT to generate a linear relationship between the RED and the weighted subtraction of a high and low energy scan of a phantom. The purpose of this study was to investigate the applicability and robustness of this method to linac-based cone beam CT (CBCT) imaging with the aim of facilitating dose of the day calculations in future. Material and Methods Dual energy (DE) scans of a Catphan 503 and a CIRS 062MQA phantom with the electron density inserts positioned at the isocenter plane were acquired (two consecutive scans; 70 kVp (LE) and 130 kVp (HE); Elekta XVI 5.0 equipped with an anti-scatter grid [2]). The CIRS was scanned in standard (CIRS M) and small configuration (CIRS S) to examine the effect of the different scatter conditions. The HE and LE scans were combined according to ΔHU=(1+α) HU HE – α HU LE [1]. α was determined by maximizing the value of the R 2 -parameter of the ΔHU-RED linear fit. To characterize the robustness of this DE method, α and the fit parameters from one phantom were applied in turn to the images of the other two phantoms and the absolute difference in RED with the fitted curve was calculated. For comparison, a bilinear interpolation was performed for a 120 kVp single- energy (SE) reference scan. Results Excellent goodness of fits (R 2 >0.99) were observed for both SE and DE and all phantoms (Fig. 1). Consequently, small differences between estimated and specified RED

Changes in D95%-PTV (above) and healthy liver NTCP (below), as evaluated on the contours used for treatment planning when compared to all other contours sets, versus the DSC of the corresponding pairs of PTVs (above) and liver contours (below). Conclusion The dosimetric impact of delineation variation was significant, and may have substantial impact on toxicity and tumor control. Adequate tumor and OAR delineation is essential, especially in high precision SBRT, where small CTV-PTV margins are used. Dice similarity coefficients did poorly predict the dosimetric impact of delineation uncertainties. EP-2145 Understanding variation in CT radiomics features – a potential method to reduce feature space L.C. Holloway 1,2,3,4,5 , C. Brink 2 , M. Field 6 1 Ingham Institute and Liverpool and Macarthur Cancer Therapy Centres, Radiation Oncology, Sydney, Australia 2 University of Southern Denmark, Institute of Clinical Research, Odense, Denmark 3 University of Wollongong, Centre for Medical Radiation Physics, Wollongong, Australia 4 University of Sydney, School of Physics, Sydney, Australia 5 University of New South Wales, school of medicine, Sydney, Australia 6 University of New South Wales and Ingham Institute, school of medicine, Sydney, Australia Purpose or Objective Radiomics features, as determined from medical images are increasingly being used in developing predictive models. Given the large number of features a reduction of parameter space is desirable. There are many radiomics features which may be impacted by a number of institution dependent factors such as, scanner type, scanner settings, image resolution, signal intensity bin resolution and contours. The institution dependent factors could change the feature value and even the ranking among feature values for specific patients; and should be considered carefully for cross institution decision models. This study investigates the feature instability as a method to reduce the feature space. Furthermore it is demonstrated that large correlations among the features can be used to reduce the number of Twenty non-small cell lung cancer patient CT datasets complete with radiotherapy gross tumour volumes (GTVs) were used as the baseline dataset. For each of these images we firstly resampled the voxel size such that 8 original voxels were merged into a single voxel and also split a single voxel into 8 smaller voxels. Secondly, the signal intensity values were binned into units of 25 Hounsfeld units and 10 Hounsfeld units. Thirdly, contours were expanded and contracted by one voxel from the original contour. For all datasets a set of 79 radiomics features calculated based on the original set of features presented by Aerts et al were determined. A Pearson initial model features. Material and Methods