ESTRO 2020 Abstract book

S906 ESTRO 2020

included cardiac dose-volume receiving ≥10 Gy(heart V10), mean lung dose, baseline hemoglobin(b-Hb) level, h-NLR and total cisplatin dose before CCRT finished(t-cisplatin). On multivariate LR, heart V10 (>86.5 vs. ≤86.5%; OR 0.278; CI 0.079–0.976, p = 0.046), b-Hb level (>11.25 vs. ≤11.25g/dL; OR 11.536; CI 2.036–65.377, p = 0.006), h- NLR (>14.48 vs. ≤14.48; OR 0.261; CI 0.075–0.910, p = 0.035) and t-cisplatin (>147.5 vs. ≤147.5 mg/m2; OR 4.966; CI 1.446–17.051, p = 0.011) remained significantly associations with response. Heart V10 also had a trend to OS on univariate analysis(HR 1.643; CI 0.994–2.715, p = 0.053). Further relationship between h-NLR and dosimetric parameters showed statistically significant correlation with heart V10(>94.5% vs. ≤94.5%, Spearman's correlation coefficient r = 0.259, P = 0.018). Conclusion Heart V10, b-Hb level, h-NLR, and t-cisplatin were predictors for CCRT response. The h-NLR also correlated with heart V10, as the only dosimetric parameters related to the severity of treatment-elicited inflammation. Radiotherapy technique improvement, like proton therapy, to spare the volume of low dose cardiac irradiation, maybe a choice to reduce the severity of inflammation and to impact the treatment response. PO-1579 Deep learning based gross tumor volume definition on planning CTs of soft tissue sarcoma D. Lang 1,2,3 , J.C. Peeken 1,3,4 , M.B. Spraker 5 , M.J. Nyflot 5,6 , S.E. Combs 1,3,4 , J.J. Wilkens 2,3 , S. Bartzsch 1,3 1 Helmholtz Zentrum München, Institute of Radiation Medicine, Munich, Germany ; 2 Technical University of Munich, Physics Department, Munich, Germany ; 3 School of Medicine and Klinikum rechts der Isar- Technical University of Munich TUM, Department of Radiation Oncology, Munich, Germany ; 4 Deutsches Konsortium für Translationale Krebsforschung DKTK, Partner Site Munich, Munich, Germany ; 5 University of Washington, Department of Radiation Oncology, Seattle, USA ; 6 University of Washington, Department of Radiology- Seattle, Seattle, USA Purpose or Objective Medical imaging data represents the main information source of radiation therapy treatment planning. In recent years deep learning has been introduced as a very powerful machine learning technique that proved particularly useful in image classification and segmentation. In comparison to conventional radiomics, which relies on handcrafted feature extraction, deep learning allows for end-to-end training that can be applied to more advanced tasks such as tumor segmentation. We studied the ability of deep learning for automated segmentation of gross tumor volumes (GTV) of soft tissue sarcomas (STS) based on CT imaging. Material and Methods The data set used consisted of two independent patient cohorts wit 80 patients from the Technical University of Munich (TUM) and 87 patients from the University of Washington/Seattle Cancer Care Alliance (UW). Both cohorts involved a very diverse set of samples with STS located in different parts of the body, leading to a large variance in tumor size. Segmentation of the GTV was manually done by a radiation oncologists at both institutions. The TUM cohort was split into training and validation sets with 64 and 16 patients and the UW cohort was used as testing set. A fully convolutional 2.5D network based on densely connected building blocks was set up, following the strategy of Jégou et al. (arXiv:1611.09326). In contrast to the paper, the dimension reduction in the downsampling path was not achieved by max pooling but by a 4x4 filter size with a stride of 2x2 and zero padding in the convolutional layers. For training, the CTs were cropped in the axial plane resulting in a size of 256x256 pixels. In longitudinal direction 3 consecutive slices were used. The model consisted of two dense blocks in the

Conclusion The developed MRI-radiomics signature can be a valuable tool for the prognosis of LANPC pts in non endemic area. It could be integrated into a multidimensional nomogram including disease stage at diagnosis and baseline EBV-DNA plasma load. This might define pts’ prognosis more accurately and lead to the development of tailored treatment through radiomics, biological and clinical characteristics. PO-1578 Heart V10 is related to treatment-elicited inflammation and clinical response in esophageal cancer. Y. Ho 1 , J. Lin 1 , M. Ko 1 , T. Chou 1 , L. Hung 1 , C. Huang 1 , T. Chang 1 1 Changhua Christian Hospital, Department of Radiation Oncology, Changhua, Taiwan Purpose or Objective For esophageal cancer with cervical location or non- surgical candidate, definitive concurrent chemoradiation( CCRT ) is a standard treatment. Initial treatment response may correlate with survival, but, few validated markers are available to predict . Recent studies also correlated the treatment-elicited inflammation with tumor progression and survival. The purpose of this study is to evaluate clinical variables, chemotherapy dose, and normal tissue parameters associated with response. Material and Methods From 2010 to April 2016, 86 patients with esophageal squamous cell carcinoma finished CCRT at our institution. Patients' clinical, dosimetric, and laboratory data at baseline and during-CCRT were collected by review of medical records. Cox regression was used to assess survival. ROC curve with the Youden index was chosen as thresholds to increase specificity and to dichotomize continuous variables for response and highest neutrophil to lymphocyte ratio during-CCRT(h-NLR) . The response was defined on RECIST 1.1 by the first post-CCRT computed tomography scan follow-up. Responders defined as the sum of complete and partial responses and non-responders as the sum of stable and progressive diseases. Logistic regression(LR) was used to evaluate those variables associated with response. Spearman's rho test was used to show the relationship between h-NLR and dosimetric parameters. Results The mean age was 59.7 years, male predominance(88%), with 29% stage II(n=25) and 71% stage III(n=61). The majority received cisplatin-based CCRT(97%). Univariate analysis of responders(n = 50) vs. non-responders(n = 36) showed significant difference in median OS at 10 vs. 22 months and 2-year OS of 8.3% vs. 46%(HR 0.316; CI 0.195– 0.511, p < 0.001). Predictors of response on univariate LR