Abstract Book

S1149

ESTRO 37

Electronic Poster: Physics track: (Quantitative) functional and biological imaging

EP-2091 Complementarity of PET-texture features with respect to information given by conventional indexes. M. Carles 1 , T. Bach 2 , I. Torres 3 , D. Baltas 1 , U. Nestle 4 , L. Martí 3 1 Medical Center- Faculty of Medicine- University of Freiburg, Division of Medical Physics- Department of Radiation Oncology, Freiburg, Germany 2 Medical Center- Faculty of Medicine- University of Freiburg, Administrative and Clinical Informatics- Department of Radiation Oncology, Freiburg, Germany 3 Hospital Universitario y Politécnico La Fe, Clinical Area of Medical Imaging, Valencia, Spain 4 Medical Center- Faculty of Medicine- University of Freiburg, Department of Radiation Oncology, Freiburg, Germany Purpose or Objective The interest in the quantification of tumor heterogeneity based on texture features (TF) with PET has recently increased. However the complementary information that PET-TF could provide with respect to the conventional indexes (SUV max , metabolic volume V and TLG) has been shown to be dependent of the method employed to resample the tumor voxel intensities, which is a preprocessing required for TF computation. The aim of our work was to study the complementarity of TF and conventional indexes information when two resampling methods were applied for TF computation. Material and Methods Thirty-one lung patients (36 lesions) were retrospectively analyzed. The first method (RN) resampled the tumor intensities with a constant number of bins (N) along all the 36 lesions: N=16, N=32 and N=64. For the second method (RW), the width of the resampling bin (SUV resolution, W) was constant along the lesions: W= 0.05, W=0.1 and W=0.5. Complementarity was evaluated in terms of Spearman's correlations (p<0.05). We considered TF with added value the ones that did not show strong correlation with any of the 3 conventional indexes, i. e. p>0.05 or (p<0.05, r<0.8) for all indexes. Results Results showed that TF complementarity with respect to conventional indexes was dependent on the resampling methods evaluated. For RW stronger correlations with SUV max were found, figure 1.a. TF with RN were more correlated with volume, figure 1.b. RW gave rise to a larger number of TF with added value, table 1.

Conclusion From the results, RW with a SUV resolution of 0.1 should be recommended as the resampling method to use in PET-TF computation for tumor heterogeneity quantification. EP-2092 Generative Model of Functional RT-Plan Chest CT for Lung Cancer Patients Using Machine Learning B.S. Jang 1 , J.Y. Chang 2 , A.J. Park 3 , H.G. Wu 1 1 Seoul National University Hospital, Radiation Oncology, Seoul, Korea Republic of 2 SMG-SNU Boramae Medical Center, Radiation Oncology, Seoul, Korea Republic of 3 SELVAS AI Incorporation, Artificial Intelligence Research and Development Laboratory, Seoul, Korea Republic of Purpose or Objective Functional image-guided radiotherapy (RT) planning for normal lung avoidance has recently been introduced. SPECT/CT can help identify the functional area of patients, but it is associated with delayed treatment time, additional costs, and unexpected radiation exposure. In this study, we propose a machine learning algorithm that can generate functional chest CT images based on conditional generative adversarial networks (cGAN) using publicly available ‘pix2pix’ implementation. Material and Methods We collected a total of 54 lung perfusion SPECT/CT image sets from lung cancer patients who had been treated at the Seoul National University Hospital. Each SPECT and CT set consists of 161 images. The CT-to-SPECT image pair where there was blank, was outside from the lung parenchymal field, or was not matched anatomically due to the patient's breath was removed under physician’s judgment. After we excluded inappropriate images, 3054 CT-to-SPECT image pair of training sets (49 patients) and 400 of test sets (5 patients) were selected, respectively. Model was trained using ‘pix2pix’ implementation. Results The model was evaluated based on the Multiscale SSIM (MS-SSIM). With 400 image pair of test set, we obtained a lung SPECT/CT fusion image demonstrating MS-SSIM of 0.87 (0.60 – 0.99) compared to original image. This result showed that it could generate functional areas of lung parenchyma directly from chest CT using machine learning algorithm.It suggest that our trained model using cGAN can generate functional areas from RT-plan chest CT images. This could be used for functional image- guided RT planning, sparing patient’s lung function without additional imaging modalities and costs. Further work including planning studies entails with much more training and test set. Conclusion It suggests that our trained model using cGAN can generate functional areas from RT-plan chest CT images. This could be used for functional image-guided RT planning, sparing patient’s lung function without additional imaging modalities and costs. Further work including planning studies entails with much more training and test set. EP-2093 Dose–function histogram evaluation using 99mTc-GSA SPECT/CT images for SBRT planning for HCC R. Toya 1 , T. Saito 1 , S. Shiraishi 2 , Y. Kai 3 , R. Murakami 4 , T. Matsuyama 1 , T. Watakabe 1 , F. Sakamoto 2 , N. Tsuda 2 , Y. Shimohigashi 3 , Y. Yamashita 2 , N. Oya 1