Abstract Book

S1176

ESTRO 37

Purpose or Objective To explore the possibility of texture features from daily megavoltage computed tomography (MVCT) images on predicting radiotherapy outcome of non-small cell lung cancer (NSCLC). Material and Methods Daily MVCT images from 18 patients with peripheral primary non-small cell lung cancer (histo-pathological stage T1N0M0) were prospectively analyzed. Craniocaudal respiratory movement observed under 2D simulator should be less than 1cm as enrollment criteria. MVCT images were obtained covering the whole gross tumor volume (GTV) plus 2cm extension in superior-inferior direction. Ten patients were treated with radiotherapy prescription dose of 4.0~7.0Gy/fraction, 7 patients with 10.0~11.0Gy/fraction and 1 patient with 15.0Gy/fraction. MVCT image scanning parameters: acquisition pitch 4mm/r and reconstruction interval 1mm. Image resolution is 512x512 pixel. Gross tumor volumes (GTV) were delineated by 3DSlicer with lung window setting (window width, 1500 HU; window level, -700 HU). Texture features of GTVs were extracted by the Imaging Biomarker Explorer (IBEX) software. To test whether the texture features changed with stable and consistent variation trends during the treatment course, each texture feature is observed and their values were The average GTV volume of 18 patients was 12.0cc. Artifacts on MVCT images from the respiratory movement did not seriously affect the delineations. Fifty-two texture features from 5 categories were extracted from each set of MVCT images, which features were selected to cover the diverse range of features that have been used in previous texture feature studies using CT images of NSCLC. Forty (77%) texture features presented stable and consistent variation trends during the treatment course. Two texture features: Meanabsolutedeviation, Graylevelnonuniformity presented stable and consistent variation trends on MVCT images from 5 and 4 patients respectively. Texture features InverseVariance, LongRunEmphasis, RunPercentage, Coarseness, GlobalMean, GlobalStd and Variance presented stable and consistent variation trends on MVCT images from 3 patients. Conclusion Some of the texture features extracted from the daily MVCT images obtained by TomoTerapy have a steady and consistent variation trend during the radiotherapy course, which may be used to predict radiotherapy outcomes in patients with NSCLC. Further investigations are expected to relate the texture features from TomoTherapy MVCT and radiotherapy outcomes. EP-2135 Statistical motion masks to identify sliding surfaces for motion models used on an MR-Linac B. Eiben 1 , E.H. Tran 1 , M.J. Menten 2 , A. Wetscherek 2 , D.J. Hawkes 1 , U. Oelfke 2 , J.R. McClelland 1 1 University College London, Medical Physics and Biomedical Engineering, London, United Kingdom 2 The Institute of Cancer Research and The Royal Marsden NHS Foundation Trust, Joint Department of Physics, London, United Kingdom Purpose or Objective Respiratory motion is a challenge for precise radiotherapy delivery for lung cancer patients. MR-Linac technology facilitates real-time monitoring of lung tumour motion using 2D cine-MR acquisitions. However, motion of the full 3D patient anatomy including the tumour and organs at risk remains, due to relatively slow 3D MR acquisition and reconstruction techniques, an unsolved problem. In this regard motion models are a promising tool to estimate the full 3D motion of the patient during treatment based on a surrogate signal. Building such plotted. Results

models requires motion measurements obtained from image registration. However, respiration causes the lung and more inferior structures, such as the liver to slide along the pleural wall. Sliding is challenging for registration methods which usually regularise motion estimates to be smooth. Hence, in order to efficiently deal with this type of motion, sliding surfaces need to be identified. We present a method that first builds a statistical shape model (SSM) from motion masks, which include the lung and inferior sliding structures, based on 4DCT images. Those shape models can then be fit to a variety of image modalities - including MRI - and different resolutions to delineate sliding surfaces. Material and Methods Mid-position images were created by group-wise image registration for 32 lung 4DCT image sets. From each mid- position image a motion mask was calculated which includes organs enclosed by the sliding surface such as the lungs, mediastinum, diaphragm, and liver (Figure 1). In order to measure the inter-patient variation of the motion masks, these masks need to be aligned. This was done by group-wise registration of the mid-position images and then applying the transformations to each corresponding mask. A population average mask was created and then transformed into a mesh representation. An SSM was built from the average mesh and transformed back to each individual mid-position image where the intensity gradients along the mesh normals were calculated for each node. The gradient information is then used to fit the model to an image.

Results The performance of the algorithm was evaluated quantitatively using a leave-one-out strategy on the 32 4DCT images. The statistical motion mask with ten modes of variation was fit to the left-out image and then compared to its original mask. The mean Dice coefficient and mean contour distance were calculated as 0.96±0.03 and 3.8±2.8mm, respectively. Furthermore, the algorithm was applied to the first level of a multi-resolution, motion-model based image reconstruction from MR images. The visual assessment shows good positioning of the mask in the low resolution image. For an example of the leave-one-out fitting to a CT dataset and the low- resolution MR image see Figure 2.