ESTRO 2021 Abstract Book

S718

ESTRO 2021

significant.

Conclusion We identified 2 radiomic features, making it possible to identify a priori patientswho will present a lower risk of LR-PFS after a CT-RT. Those pts could potentially benefit of exclusive CT-RT, thus avoiding the risks of surgical procedures. PD-0881 Deep learning dose prediction as a tool to evaluate treatment complications based on NTCP models C. Draguet 1,2 , A. Barragan-Montero 1 , S. Michiels 1 , G. Defraene 2 , M. Thomas 3,4 , K. Haustermans 2,4 , J. Lee 1 , E. Sterpin 1,5 1 UCLouvain, Institut de recherche expérimentale et clinique, Molecular Imaging Radiotherapy and Oncology , Brussels, Belgium; 2 KULeuven, Department of Oncology, Laboratory of experimental radiotherapy, Leuven, Belgium; 3 KULeuven, Department of Oncolocy, Laboratory of experimental radiotherapy, Leuven, Belgium; 4 UZLeuven, Department of radiation oncology, Leuven, Belgium; 5 KULeuven, Department of oncology, Laboratory of experimental radiotherapy, Leuven, Belgium Purpose or Objective Dose prediction models based on deep learning (DL) algorithms become a common tool for automatic radiotherapy treatment planning. However, the combination of these algorithms with normal tissue complication probability (NTCP) models has the potential to automatically detect patients at high risk of toxicity, and thus, to trigger the evaluation of alternative treatments to reduce such complications, without the need of spending countless hours planning treatments manually. This work aims to evaluate the accuracy of our DL dose prediction model for intensity-modulated radiation therapy (IMRT) treatments in combination with NTCP models to detect esophageal cancer patients at high risk of toxicity. Materials and Methods Our DL model, a UNet architecture with dense connections, was trained with a database of 60 patients and tested on 9 patients. For the test patients, the predicted and ground truth 3D dose distributions were used to extract relevant dose-volume metrics that were the input for NTCP models. Two NTCP models were used to evaluate: 1) postoperative pulmonary complications such as pneumonia, respiratory failure and respiratory distress syndrome [NTCP p , Thomas et al. 2019], and 2) cardiac complications (pericardial effusion) [NTCP c , Beukema et al. 2020]. NTCP p involved the mean lung dose, age, histology type and body mass index as predicting variables, while NTCP c involved the mean heart dose only. The absolute error for the two considered NTCP models for the predicted and ground truth doses were compared. Moreover, in order to evaluate the impact of the dose prediction error on the NTCP values as a decision support tool to select patients for alternative treatments, a proton (PT) plan was created for each test patient. The accuracy of our DL model for patient referral is then evaluated based on ΔNTCP thresholds with respect to the PT plans. According to the Dutch Society for Radiotherapy and Oncology, a ∑ΔNTCP ≥ 15 % for grade 2 complications is a necessary condition for redirecting patient to PT. Results The MAE for the metrics accounted in the NTCPs was 0.79 ± 0.56 Gy for the mean lung dose (MLD) and 1.57 ± 1.45 Gy for the mean heart dose (MHD). The MAE for the predicted complications was 1.93 ± 1.52 % for NTCP p and 2.56 ± 2.39 % for NTCP c . Only two patients presented an error above 5% (patient 5, 5.57% for NTCP p and patient 1, 7.6% for NTCP c ). The comparison with the PT plans led to the same clinical decision (referral to PT) for both predicted and ground truth doses.