ESTRO 37 Abstract book

ESTRO 37

S535

clinicians on the CT images (512 x 512 pixels, pixel size 0.9766 ). Two quantisation methods were studied: Group Uniform Quantiser (GUQ) and Individualised Uniform Quantiser (IUQ). GUQ has an intensity range optimised for the whole patient group whereas IUQ quantises an intensity range optimised for each individual patient. Both methods were applied to each patient dataset in levels from 4 to 256 to understand the impact on discriminating powers of entropy calculated from 3D GLCM, one of the most robust features reported. Results Figure (1a, b) plots entropy vs quantisation levels for all cases using GUQ and IUQ. Intersections in the entropy lines are present in both methods, which emphasises the instability and sensitivity of entropy to choice of quantisation levels. Entropy was ranked from lowest to highest for all cases in all quantisation levels. The ranks do not correspond with GTV size, indicating entropy is not a function of tumour size. Cases were grouped by entropy into Group Low (GL - lowest 19) and Group High (GH – highest 18). For a given number of bins, only 2/3 cases stay in the same group (GH or GL) when one quantisation method is compared with the other. With GUQ, 10 cases remained in GL whereas 7 remained in GL with IUQ (3 remain unchanged with both methods). 9 cases remained in GH for GUQ and 10 cases for IUQ (6 cases unchanged for both methods) for all quantisation levels. Table 1 shows large changes from 8 to 16 bins and 32 to 64 bins for both methods, stability increased with GUQ from 64 bins but less so for IUQ. Therefore using quantisation levels from one single patient scan may not produce a robust result for grouping studies.

Conclusion In general, a large inter-center variability in implementation of CT scans, image reconstruction and especially in CT-to-SPR conversion was found. The benefit of a future standardization is obvious: It would reduce the time-intensive site-specific efforts as well as variations in treatment quality. Due to the interdependency of multiple parameters, no conclusion can be drawn on the derived SPR accuracy and its inter- center variability. As a next step within the EPTN, an inter-center comparison of CT-based SPR prediction accuracy will be performed with a ground-truth phantom. H.Y.C. Wang 1 , E. Donovan 1 , A. Nisbet 2 , S. Alobaidli 3 , I. Phillips 4 , V. Ezhil 4 , P. Webster 5 , M. Ferreira 5 , P. Evans 1 1 University of Surrey, Centre of Vision- Speech and Signal Processing, Guildford, United Kingdom 2 Royal Surrey County Hospital, Medical Physics, Guildford, United Kingdom 3 Kuwait Cancer Control Centre, Radiotherapy Physics, Kuwait, Kuwait 4 Royal Surrey County Hospital, St Lukes Cancer Cantre, Guildford, United Kingdom 5 Alliance Medical Limited, London, United Kingdom Purpose or Objective Grouping of patients based on a threshold value for underlying texture features extracted from standard CT scans is commonly used in prognostic predictive studies including radiomics. Various quantisation levels to reduce noise and artefacts in analyses have been used without a firm recommendation. The aim of this study is to assess the effect of quantisation levels on image entropy (intra- voxel heterogeneity) from Grey Level Co-occurrence Matrix (GLCM) for lung cancer patients based on two methods with different intensity range to arrive at a robust methodology for standardisation and inter-cohort comparison. Material and Methods 37 NSCLC patients with Positon Emission and Computed Tomography (PET/CT) datasets were included. Gross tumour volumes (GTV) were manually outlined by PO-0970 Robustness of Texture as a Biomarker in Radiomics Applications

Figure 1: Entropy for All Cases with GUQ and IUQ

Table 1: Patients Changed Group when n Changes by 1

Conclusion Entropy calculated from GLCM is extremely sensitive to quantisation levels and method, but stable to changes in size of the tumour volume. We recommend using GUQ with at least 64 bins to give robust results for inter- cohort comparison. Acknowledgement: Alliance Medical Limited for funding H Wang.