Abstract Book

S1161

ESTRO 37

5 Université catholique de Louvain, IREC/EPID, Brussels, Belgium Purpose or Objective PET-guided dose painting (DP) aims at improving the efficacy of radiation therapy by focusing the treatment on radioresistant tumour regions identified through PET imaging. Importantly, this strategy depends on the spatial accuracy of the images which are used to guide the treatment. However, PET images are blurred as a result of the low spatial resolution of PET imaging systems, which may lead to inaccurate biological mapping of the tumour. In this context, we evaluated the performance of a novel PET image deblurring algorithm. Material and Methods We retrieved the database from Christian et al. [1], containing FDG autoradiographs of 14 mice tumours and the corresponding registered PET images. Next, we applied a novel iterative deblurring algorithm to these PET images. This algorithm involves a positivity constraint and local dynamic constraints on the PET signal admissible range, to avoid artifacts and noise amplification. Finally, we analysed whether image deblurring allowed to increase the similarity between PET images and the corresponding autoradiographs, using two indexes: the Spearman’s correlation coefficient (r S ) and the structural similarity index (SSIM). The SSIM is a novel index which tries to mimic the human visual perception [2]. Results PET image deblurring increased the visual impression of similarity between autoradiographs and PET images (as illustrated in the attached figure). Quantitatively, deblurring increased the mean r S between autoradiography and PET images from 0.787 to 0.795 (p = 0.44), and the mean SSIM from 0.702 to 0.756 (p = 0.035).

Conclusion The considered deblurring algorithm improved the spatial accuracy of PET images. This constitutes a step towards better biological mapping of the tumour in the context of PET-guided DP. 1. Christian N, Lee JA, Bol A, et al. The limitation of PET imaging for biological adaptive-IMRT assessed in animal models. Radiother Oncol. 2009;91:101-6. 2. Sampat MP, Wang Z, Gupta S, et al. Complex wavelet structural similarity: a new image similarity index. IEEE Trans Image Process. 2009;18:2385-401. EP-2112 How accurate should a GTV delineation be for radiomics? A study in NSCLC patients J.E. Van Timmeren 1 , R.T.H. Leijenaar 1 , W. Van Elmpt 2 , P. Lambin 1 1 The D-Lab: Decision Support for Precision Medicine, GROW - School for Oncology and Developmental Biology- Maastricht University Medical Center, Maastricht, The Netherlands 2 Department of Radiation Oncology MAASTRO, GROW - School for Oncology and Developmental Biology- Maastricht University Medical Center, Maastricht, The Netherlands Purpose or Objective Radiomics shows high potential for performing outcome predictions for non-small cell lung cancer (NSCLC) patients. However, delineations of the region of interest (ROI) are prone to interobserver variability which limits the reproducibility and robustness. Moreover, the introduction of new concepts based on longitudinal radiomics, are hampered by the time-consuming semi- automatic delineation of the ROI. Therefore, we investigated the prognostic value of a previously validated radiomic signature after eroding the original GTV. Material and Methods A dataset of 102 NSCLC patients was used in this study to extract radiomic features from the treatment planning CT images. For each patient, the GTV was eroded using a 3D structuring element with radius’ of 3, 5, 7 and 9 mm using the open-source platform REGGUI (http://openreggui.org). Due to the small size of some patients’ GTV, the eroded contours could be calculated for 97, 87, 79 and 66 patients, respectively. The radiomic feature values of the four features of a previously published radiomic signature (Aerts et al., 2014) were calculated for each GTV. Subsequently, the coefficients of the model were rescaled based on the slope of a linear regression of the radiomic features extracted from the original GTV and the eroded GTV. Harrell’s concordance index (c-index) and calibration slope of the models were calculated. Results Linear regression of radiomic features values extracted from original and eroded contours resulted in slope values that increased linearly with the size of the 3D structuring element used to erode the contours. The slopes with corresponding coefficients of determination (R 2 ) are summarized in Figure 1.