ESTRO 37 Abstract book

S1160

ESTRO 37

EP-2105 Radiomics in response prediction to Cyberknife radiosurgery for acoustic neuroma:a pilot study I. Bossi Zanetti 1 , N.C. D'Amico 2 , E. Grossi 3 , G. Valbusa 3 , G. D'Anna 2 , D. Fazzini 2 , F. Rigiroli 4 , A. Bergantin 1 , I. Redaelli 1 , A. Maldera 1 , I. Castiglioni 5 , G. Scotti 2 , S. Papa 2 , G. Beltramo 1 1 Centro Diagnostico Italiano, Cyberknife, Milano, Italy 2 Centro Diagnostico Italiano, Imaging, Milano, Italy 3 Bracco Imaging S.p.A, Imaging, Milano, Italy 4 Università degli Studi di Milano, Postgraduation school of Radiodiagnostics, Milano, Italy 5 CNR, Istituto di Bioimmagini e Fisiologia Molecolare, Milano, Italy Purpose or Objective The objective of this study is to analyse MR images acquired before Cyberknife radiosurgery in order to predict the volumetric evolution of the tumour after the treatment. Material and Methods T1 weighted MR images of 38 patients presenting an acoustic neuroma treated with Cyberknife® at Centro Diagnostico Italiano, were acquired and analysed. These selected patients had a follow-up of at least one year (mean 4.6 years, range: 1-10 years) in order to know the tumour evolution after the treatment: 20 patients (52.6%) responded with a Volumetric Reduction (VR), 14 patients (36.8%) responded with stable volumetric evolution and 4 patients (10.5%) had a volume increase. These last two subgroups were considered as a unique set defined as patients without Volumetric Reduction (wVR). The analysed images were acquired on 1.5T machines with contrast enhanced T1-weighted sequences in the axial plane. These images were acquired before the treatment with a standardized protocol necessary as guided image for the Cyberknife® treatment. Semi- automatic tumour segmentation was carried out on MR images using the 3DSlicer image analysis software. The used editor modules is the level tracing effect, where the operator is required to define the segmented region interactively by moving the mouse over the region of interest letting the software automatically adjusting an outline where the pixels all have the same intensity value as the current selected pixel. After the tumour segmentation, the images were pre-processed, resampling label and intensity images to voxels of 1x1x1 mm. MR images features were calculated using a dedicated software developed upon the ITK framework. Shape-based, intensity-based and texture-based features were extracted. An evolutionary algorithm (a TWIST system based on KNN algorithm) was used to subdivide the dataset into training and validation test and to select features yielding the maximal amount of information. After this pre-processing, different machine learning systems(MLS) were applied to develop a predictive model based on a training-testing crossover procedure. The best neural network was a 3-layers feed forward back propagation algorithm with 8 input variables containing the maximal amount of information. Results Two training/testing groups were created (group 1: training-21; testing-17, group 2: training-17; testing-21). The neural network was used twice inverting the training/testing set. In the first analysis the sensitivity was 100%, while the specificity was 77.78%. These two results gave a global accuracy of 88.89%. In the second analysis the sensitivity was 61.54% and the specificity 100%, with a global accuracy of 80.77%. The mean value of the global accuracy was 84.83%. Conclusion The obtained results show that Machine Learning coupled with radiomics has a great potential in distinguishing, before the treatment, responders with volume reduction

from responders without volume reduction to radiosurgery. EP-2106 Simulation of oxygen and hypoxic PET-tracer distributions in 3D vascular architectures of tumours I. Paredes Cisneros 1,2,3,4 , I. Espinoza 4 , D. Nolte 5 , C. Bertoglio 5,6 , C. Karger 1,3 , A. Gago-Arias 4 1 German Cancer Research Center DKFZ, Medical Physics in Radiation Oncology, Heidelberg, Germany 2 Heidelberg University, Medical Faculty Heidelberg, Heidelberg, Germany 3 Heidelberg Institute for Radiation Oncology HIRO, National Center for Radiation Research in Oncology NCRO, Heidelberg, Germany 4 Pontificia Universidad Católica de Chile, Institute of Physics, Santiago, Chile 5 University of Chile, Center for Mathematical Modeling, Santiago, Chile 6 University of Groningen, Johann Bernoulli Institute, Groningen, The Netherlands Purpose or Objective Tumour response to ionizing radiation exhibits a strong dependence on oxygenation due to the increased radioresistance of hypoxic cells. Hypoxic PET-tracers are used to assess tumour oxygenation prior and during radiotherapy treatments in a non-invasive way. Previous studies have simulated the oxygen distribution and the resulting voxel-based time-activity curves (TAC) of the tracer in tissue sections. However, no simulations on realistic three-dimensional (3D) vascular architectures have been studied yet. This study aims to develop a tool to simulate the voxel-based oxygen and FMISO-PET dynamics for realistic 3D vascular architectures. Material and Methods Oxygen and FMISO-PET dynamic distributions were calculated by solving reaction-diffusion equations, using a 3D finite element method. 3D vascular architectures and the corresponding meshes were constructed using the open source meshing software Gmsh . In this framework, vessels were modeled as cylinders and constitute the source of oxygen (partial oxygen pressure (pO2) of 40 mmHg). For the simulation of dynamics of the FMISO distribution, an arterial input function was used. The underlying partial differential equations were solved using the open source computing platform FEniCS . The vessel geometry consisted of three perpendicular intersecting vessels, of radius 10 µm and 300 µm long. First, the stationary oxygen distribution was simulated. This distribution was then used as input to solve the time-dependent 3D FMISO distribution. The impact on FMISO dynamics of maximum oxygen consumption (gmax) and capillary permeability to FMISO (L T ) was investigated. TAC were calculated from the total FMISO concentration in the volume and the input function. Results Figure 1 shows pO2 and bound FMISO concentration (Cb) for the given geometry. For distances larger than approximately 180 µm from vessels, cells were considered necrotic and no binding of FMISO occured.

Figure 1. pO2 and Cb distributions at 90 minutes post injection.