ESTRO 2021 Abstract Book

S375

ESTRO 2021

Conclusion The findings indicate that whole saliva flow rates decrease, and xerostomia grades increase with increasing mean dose to the parotid glands. At the mean dose above 39 Gy severe xerostomia and hyposalivation persist one year of completion of RT. Dose to the parotid glands alone can explain 29-41% of the variation in salivary flow changes and 79% of the variation in xerostomia grade. Currently the sample size is small and further investigations are required. PH-0490 Deep learning predicts survival for early stage NSCLC patients treated with SBRT S. Zheng 1 , J. Guo 1 , J. Langendijk 1 , S. Both 1 , R. Veldhuis 2 , M. Oudkerk 3 , P. van Ooijen 1 , R. Wijsman 1 , N. Sijtsema 1 1 University Medical Center Groningen, Radiation Oncology, Groningen, The Netherlands; 2 University of Twente, Faculty of Electrical Engineering, Groningen, The Netherlands; 3 University of Groningen, Faculty of Medical Science, Groningen, The Netherlands Purpose or Objective Stereotactic Body Radiation therapy (SBRT) is a good alternative treatment for inoperable patients with early- stage non-small cell lung cancer (ES-NSCLC). The 2-year overall survival rate after SBRT varies between 50% and 71%. To improve patient stratification that may provide opportunities for a more personalized approach, this study aimed to develop a hybrid deep learning-based model that can predict 2-year overall survival accurately for ES-NSCLC patients by considering clinical factors and image features from pre-treatment CT- scans. Materials and Methods A training set and an independent test set A, including 189 and 81 patients who received SBRT, were prospectively collected at a medical center. External validation was performed on an independent test set B that contained 228 ES-NSCLC patients from another center. A hybrid deep learning-based model that integrated both clinical and image features was implemented for survival prediction. Image features were learned from cubic patches with the size of 64x64x64 containing lung tumors extracted from pre-treatment CT scans. Relevant clinical variables were selected from age, gender, T stage, clinical stage and biologically effective dose by univariable and multivariable analyses. The performance of the model was assessed by the area under the curve (AUC). Patients were stratified into low and high mortality risk groups by the optimal cut-off value that resulted in a maximum sum of sensitivity and specificity in the receiver operating characteristic curve on the training set. Kaplan-Meier survival curves were generated and the log-rank test was used to assess differences in the observed survival rates between two risk groups. Results Multivariable analysis showed that age and clinical stage were the prognostic factors which were also reported by other studies. Using these two clinical variables and pre-treatment CT scans, the proposed hybrid deep learning-based model achieved an AUC of 0.76 (95 CI: 0.63-0.89) and 0.64 (95 CI: 0.56-0.71) on the independent test set A and B, respectively. The Kaplan-Meier survival curves presented significant discrimination between low and high mortality risk groups on two test sets in Fig. 1. (P = 0.001, P = 0.012, respectively).