Abstract Book

S1154

ESTRO 37

EP-2099 Predicting tumour response to chemoradiotherapy in rectal cancer using dynamic contrast enhanced MRI K.M. Bakke 1 , S. Meltzer 1 , L.G. Lyckander 2 , S.H. Holmedal 3 , A. Negård 3 , K.I. Gjesdal 3 , E. Grøvik 4 , K.R. Redalen 5 1 Akershus University Hospital, Department of Oncology, Oslo, Norway 2 Akershus University Hospital, Department of Pathology, Oslo, Norway 3 Akershus University Hospital, Department of Radiology, Oslo, Norway 4 Oslo University Hospital, Department of Diagnostic Physics, Oslo, Norway 5 Norwegian University of Science and Technology, Department of Physics, Trondheim, Norway Purpose or Objective To investigate whether baseline dynamic contrast enhanced (DCE)-MRI can predict tumour regression grade (TRG) in rectal cancer patients receiving chemoradio- therapy (CRT) before surgery. This could help stratifying patients into separate treatment regimens before commencement of therapy. Material and Methods Thirty-five patients diagnosed with rectal cancer underwent DCE-MRI (T1-weighted EPI, TR = 39 ms, time resolution = 1.89 s) before pre-CRT and surgery. The images were acquired as part of a multi-echo sequence (TE = 4.6, 13.7, 22.8 ms), where the T1-weighted signal was extrapolated back to magnetization at TE = 0, to exclude T2 relaxation effects. Tumour contours were drawn by two radiologists on T2-weighted axial images. We used the extended Tofts kinetic model with a population based arterial input function, as well as non- parametric approaches for image analysis of the dynamic data, and extracted median values from the whole tumour volume. Area under curve (AUC) was determined from bolus arrival up to 90 s of scan time. Patients were divided into good responders (22 patients, TRG = 1) and poor responders (13 patients, TRG = 2-3). We used t-test and receiver operating characteristics (ROC)-curve for statistical analysis. Results K trans and AUC, both parameters reflecting a combination of blood flow and vascular permeability, were significantly different between the two groups (p-value < 0.05 and < 0.01 respectively, for both radiologists), with lower values indicating favourable response. ROC-curve analysis showed a sensitivity and specificity of 73 % and 61 % for K trans (p-value = 0.04) and 91 % and 61 % for AUC (p-value = 0.01). Both parameters were normally distributed with no outliers. Conclusion We show that parameters derived from DCE-MRI can be used to predict TRG already at time of diagnosis. The AUC parameter was slightly better than K trans at predicting poor tumour response. Since AUC is more easily obtained than K trans which requires more modelling and an arterial input function, it can be more easily automated in a clinical setting. EP-2100 Development of a filter-based method for multicenter PET image harmonization in radiomic studies S. Reuzé 1,2,3 , F. Orlhac 4 , M. Ricard 5 , D. Vallot 6 , W. Ksouri 7 , A. Laprie 8,9 , L. Dercle 10,11 , E. Deutsch 2,3,12 , C. Robert 1,2,3 1 Gustave Roussy, Radiotherapy Department- Medical Physics Unit, Villejuif, France 2 Paris-Saclay University, Faculty of Medicine, Le Kremlin- Bicêtre, France 3 INSERM U1030, Molecular Radiotherapy, Villejuif, France 4 CEA-SHFJ, IMIV, Orsay, France 5 Gustave Roussy, Nuclear Medicine Department- Medical

Physics Unit, Villejuif, France 6 Institut Universitaire du Cancer de Toulouse Oncopole, Nuclear Medicine Department, Toulouse, France 7 Hôpital Privé d'Antony, Nuclear Medicine Department, Antony, France 8 Institut Universitaire du Cancer de Toulouse Oncopole, Radiotherapy Department, Toulouse, France 9 INSERM U1214, Toulouse Neuro Imaging Center- Toulouse University, Toulouse, France 10 Gustave Roussy, Nuclear Medicine Department, Villejuif, France 11 INSERM, U1015, Villejuif, France 12 Gustave Roussy, Radiotherapy Department, Villejuif, France Purpose or Objective The role of quantitative PET imaging for predicting treatment outcome has been widely proven in the literature. A better knowledge of physical and physiological processes combined with more computing power lead to large improvements in acquisition/ reconstruction methods allowing a better precision of images. Despite a better image quality, they introduced strong biases in image quantization. Several accreditation protocols have been developed for standardized uptake value quantization, but are possibly not adapted to texture analysis. The aim of this study was to develop and validate a post-processing method permitting a standardization of retrospective PET images before radiomic analysis. Material and Methods A homogeneous Jaszczak and a Triple Line phantom (3 FDG-filled tubes, 1mm diameter) were acquired on 2 devices (PET1, PET2). 22 volumes of interest (VOI) were drawn in the Jaszczak on 9 acquisition/reconstruction protocols (PET1: N=8, PET2: N=1). An in-home Python code was developed for spatial resampling of images to a new voxel size of 2x2x2mm 3 . Spatial resolution was calculated on Triple Line and the set with the worse spatial resolution was chosen as reference. A gaussian filtering was applied to all other images to reach the reference spatial resolution (Fig.1). A multicenter clinical validation of the method was performed retrospectively in 4 institutions totalizing 137 patients, divided in 5 protocols on 3 PET devices: PET1.1, PET1.2, PET2.1, PET2.2, PET3 with respectively 29, 24, 24, 32, 28 patients. In each patient, we segmented a homogeneous hepatic VOI. Process of regridding, filtering and extraction of 39 radiomic features was applied both on phantom and patient data. A robust (non-variable) feature was defined by a non-significant difference between couples of devices according to Wilcoxon’s tests (p(feature PETa , feature PETb )>0.05).

Results Spatial resolution of reference dataset evaluated on phantom acquisitions was 7.9mm. Before standardization, only histogram features were robust while varying PSF,