Abstract Book

S748

ESTRO 37

91.9%, without significant difference between NSCLC and metastases (p=0.728) or central and peripheral lesions (p=0.511). Marginal recurrence rate was 3.7%, local progression free survival (PFS) and overall survival (OS) were 53.3% and 62.2%, respectively. >=G3-toxicity was <4%, except dyspnea: with baseline >=G3-dyspnea of 6% there was a maximum increase of 8.5% 2 years after SBRT. There was no significant difference in newly appearing >=G3 toxicity for central and peripheral tumors. Low differentiation status (p=0.034), high SUVmax in pre-therapeutic FDG-PET-scans (p<0.001) and lower prescription dose (p=0.043) predicted for worse local control. Patients with metastases (p=0.027) (especially when deriving from gastro intestinal tumors, p < 0.001) showed lower PFS. Higher age (p=0.032) and T- stage (for NSCLC patients) (p=0.023) were predictive for lower OS, the latter also for local PFS (p=0.050). Patients with lower initial FEV1 had an increased risk of marginal recurrence (p< 0.001). Data on QoL is under evaluation. Conclusion This prospective trial confirms high local control and moderate toxicity of SBRT for elderly and comorbid patients treated with moderate-dose SBRT for pulmonary lesions. EP-1370 Generation of tumor-regression CT during definitive CCRT of the lung cancer using deep learning K.S. Kim 1 , M.Y. Kim 1 , H.Y. Kim 2 , C.W. Choi 1 , K.M. Yang 1 1 Dongnam Institute of Radiological and Medical Sciences, Radiation Oncology, Busan, Korea Republic of 2 Artificial Intelligence Research Institute, Research and Developmet, Seongnamsi, Korea Republic of Purpose or Objective For the utilization of adaptive lung radiotherapy, we generated predicted tumor-regression computed tomography (CT) image using conditional generative The CT images of twenty patients who were treated with definitive concurrent chemo-radiotherapy of the lung cancer were used. Each patient had baseline CT (before treatment) and evaluation CT at radiation dose 40Gy (target CT). These pairs of CT images were aligned using MIM software and tumor-containing slices were used. Training set consisted of 14 patients with 391 images and 6 patients with 154 images were included in the test set. cGAN model can be trained to learn a pix-to-pix mapping to predict tumor-regression CT from baseline CT and target CT using generator networks and discriminator networks. Accuracy of the predicted tumor-regression CT were evaluated against target CT using dice similarity score (DSC) of the tumor lesions and volume of the Training process took 5 hours. Predicted tumor-regression CT showed decreased tumor volume while keeping chest wall, body structures and spinal cord. Tumor volume was reduced to 53 ± 10 % in target CT and 48 ± 20 % in predicted tumor-regression CT. DSC of the tumor volume was 0.61 ± 0.17. DSC of the large tumor was greater than that of the small tumor. adversarial nets (cGAN). Material and Methods tumors. Results

planning or to calculate dose distribution to the regressed tumor in advance. More large data and well designed train set configuration is essential to elevate the accuracy of the predicted tumor-regression CT images. EP-1371 Stereotactic body radiotherapy for stage I lung cancer. P.M. Samper Ots 1 , C. Vallejo Ocaña 2 , P. Alcantara 3 , M.D.M. Puertas 4 , M. Rico Osés 5 , M.L. Couselo 6 , A. Sotoca Ruiz 7 , J. Luna Tirado 8 , J.L. Monroy 9 , P. Almendros Blanco 10 1 Hospital Rey Juan Carlos, Servicio de Oncologia Radioterapia, Mostoles - Madrid, Spain 2 Hospital Ramon y Cajal, Servicio de Oncología Radioterapia, Madrid, Spain 3 Hospital Clinico, Servicio de Oncología Radioterapica, Madrid, Spain 4 Hospital Miguel Servet, Servicio de Oncología Radioterapica, Zaragoza, Spain 5 Complejo Hospitalario de Navarra, Servicio de Oncologia Radioterápica, Navarra, Spain 6 Hospital Central de la Defensa, Servicio de Oncología Radioterápica, Madrid, Spain 7 Hospital Ruber Internacional, Servicio de Oncología Radioterápica, Madrid, Spain 8 Fundación Jimenez Diaz, Servicio de Oncología Radioterápica, Madrid, Spain 9 Hospital de Alzira, Servicio de Oncología Radioterápica, Valencia, Spain 10 Hospital General de Valencia, Servicio de Oncología Radioterápica, Valencia, Spain Purpose or Objective The purpose of this study was to evaluate treatment patterns and outcomes of stereotactic body radiotherapy (SBRT) for early primary non-small cell lung cancer (NSCLC). Material and Methods Multicentric retrospective study of 139 patients with early non-small cell lung cancer, treated in ten Spanish centers between 2010 and 2016, 124 (89%) were males and 15 females, with a mean age of 73.7 years, received treatment with SBRT for 148 lung lesions. 93.9% of the patients were inoperable. A total of 64 cases (43.2%) had no pathological diagnosis, 41 (27.7%) were adenocarcinoma, 38 (25.7%) epidermoid, 1 (0.7%) indiferenciated large cells and 4 (2.7%) non-small cell. 143 lesions (96.6%) had increased metabolism in the PET/CT with an average SUV of 8.5 ± 5.7. The location of the lesions was: central 44 (29.7%), medium 34 (23%) and peripheral 70 (47.3%). The mean lesion size was 22 ± 10 mm. 74 (50%) were T1a, 42 (28,4%) T1b, 26 (17,6%) T2a and 5 (3,4%) T2b. The LSI and LSD were the most frequent localization with 49 (33%) and 44 (29.7%) lesions respectively. We performed CT simulation of normal, inspiration and expiration in 123 patients (83%) and 4D CT in 25 (16.9%). A stereotaxic immobilization system was used in 137 cases (92.6%). Control of respiration system were with dampening in 137 (92.6%), tracking 7 (4.7%) and gating 4 (2.7%). The prescription dose ranges from 30 - 70 Gy, with 60 Gy being the most frequent in 120 lesions (81%). The schemes used were: 60 Gy (5 fx of 12 Gy) in 43,2%; 60 Gy (8 fx of 7.5 Gy) in 37,2%. Treatment was given on alternate days in most cases 118 (80%). BED was <100 Gy in 3 (2.1%), 100 Gy in 17 (12%) and >100 Gy in 123 (83%). Type of RT: RC3D 70 (47.3%), VMAT 52 (35.1%), IMRT 17 (11.5%) and CiberKnife 9 (6.1%). Results With a follow-up of 2 years, 68 lesions (47.6%) had complete response in CT, 63 (92.6%) of which had received a BED> 100 Gy (p = 0.003). 74 lesions (51.7%) had complete response in the PET/CT, 61 (82.4%) had received BED> 100 Gy (p = 0.008). Local failure was seen in 11 patients (7.4%), regional failure in 14 (9.5%) and

Conclusion We generated predicted tumor-regression CT images from baseline CT using deep learning algorithm. These images could be utilized to determine the timing of the adaptive