ESTRO 37 Abstract book

S1167

ESTRO 37

better results in the assessment or response than RECIST 1.1. Central liver metastases were associated to greater biliary dilation than peripheral lesions: contingency analysis indicated that the absence of biliary dilatation was more frequent in peripheral lesions than in central lesions (P=1.0x10-5). However, there was no difference between the frequency of patients who had preexisting dilation and the frequency of patients who developed it during treatment, whether the lesion was central (P = 0.330) or peripheral (P=0.522), so no significant differences were observed between SBRT and other treatments/conditions. Finally, transitory decrease in liver attenuation surrounding lesions was observed at first trimester after SBRT, when the attenuation coefficient values were lower than in any of the other revisions including baseline measurement (P<0.016) with no other significant changes during the first year of follow-up. Conclusion RECIST 1.1, the most commonly employed criteria to assess response to SBRT of liver metastases, is not the most accurate criteria for this evaluation, being more suitable CHOI criteria, and above all, PERCIST. EP-2117 Effect of Interpolation on 3D Texture Analysis of PET Imaging in Oesophageal Cancer P. Whybra 1 , K. Foley 2 , C. Parkinson 1 , J. Staffurth 3 , E. Spezi 1 1 Cardiff University, School of Engineering, Cardiff, United Kingdom 2 Cardiff University, Division of Cancer & Genetics, Cardiff, United Kingdom 3 Cardiff University, School of Medicine, Cardiff, United Kingdom Purpose or Objective Radiomics supposes that medical images can be used as quantitative, mineable data for improved tumour profiling and personalised treatment planning [1]. Three- dimensional texture analysis (TA), a subset of radiomics, is thought to reveal an abundance of quantitative information not characterised by slice-by-slice, 2D TA. Most scan reconstruction in PET/CT produce anisotropic voxel dimensions. For 3D TA, an isotropic voxel size is advised to ensure no directional bias and is achieved via interpolation. To the best of our knowledge, the optimal voxel size and approach to interpolation is not yet standardised. We explore the effect of interpolated voxel size on selected TA features for an oesophageal cancer (OC) dataset. Material and Methods A cohort of 374 biopsy-proven OC patients were included, all radiologically staged with the same scanning protocol. The feature extraction algorithms were developed in- house and validated as part of an international initiative [2]. Results were obtained from the original voxel dimension, ‘Orig’ (0.2734, 0.2734, 0.3270) mm 3 and interpolated voxel dimensions (15mm 3 , 18mm 3 , 20mm 3 , 22mm 3 , 25mm 3 ). Tumours were discretised with a fixed bin width of 0.5 SUV prior to TA. For each feature, patients were assigned a rank based on the result for the original voxel dimension (1 for lowest value). This rank was then used for a consistent comparison with 5 different linearly interpolated voxel sizes. Results Due to the quantity of metrics, we focus this report on GLCM features which have been recently shown to be repeatable [3]. Figure 1 displays the effect of interpolation on feature value and rank for selected GLCM features. We see that dissimilarity, difference entropy and correlation show systematic change and variation in rank depending on voxel dimension, whereas sum entropy does not. For the three features showing variation with interpolation there is a clear pattern of

increasing difference from the original value with reducing interpolated voxel size. Conclusion We have highlighted several texture features that vary depending on the choice of interpolated voxel size. Features sensitive to interpolation should be used with caution when comparing datasets and developing prognostic models using 3D TA metrics. A standardised approach should be adopted and future work aims to quantitatively determine feature variation and robustness. Figure 1: 4 GLCM features; Dissimilarity, Difference Entropy, Correlation and Sum Entropy. Feature values calculated for original and 5 interpolated voxel sizes. Patients ranked for each feature based on the original dimension.

References [1] R. J. Gillies et al. “Radiomics: Images Are More than Pictures, They Are Data.,” Radiology , vol. 278, no. 2, p. 151169, 2015. [2] Zwanenburg et al . Multicentre initiative for standardisation of image biomarkers, Radiother. Oncol. Vol 123 Supp 1. [3] M.-C. Desseroit et al. , “Reliability of PET/CT Shape and Heterogeneity Features…” J. Nucl. Med. , vol. 58, no. 3, pp. 406–411, 2017 EP-2118 Identifying focal point in Glioblastoma multiforme using Association rule mining. M. Jajroudi 1 , R. Reiazi 2 , A. Azarhomayoun 3 1 Shiraz university of Medical Sciences, Medical Informatics- School of Management and Medical Information Sciences-, TEHRAN, Iran Islamic Republic of 2 Medical School- Iran university of Medical Sciences, Medical Physics, tehran, Iran Islamic Republic of 3 Tehran University of Medical Sciences, Sina Trauma Research Center- Department of Neurosurgery- Sina Hospital, Tehran, Iran Islamic Republic of Purpose or Objective Glioblastoma multiforme (GBM) is the most aggressive and fatal primary brain tumor. However the prognosis is usually very poor and survival of patients is less than 1 year; subset of them have long-term survival. In fact