ESTRO 2021 Abstract Book

S561

ESTRO 2021

Model Performance The model's mean accuracy using training generator (25 runs) was 0.93 ± 0.005 (Mean Loss: 0.22 ± 0.14; AUC = 0.98). Accuracy in the test set was 0.80 (Mean Loss: 1.10; AUC = 0.83). The F1-score for predicting DIBH was 0.93 and 0.79 for training and test set, respectively. The regions over the cardiac shadow, central mediastinum, and left upper lobe of the lung were most relevant for DIBH prediction, as viewed on back-projected activation maps.

Conclusion Deep learning methods can predict suitability for DIBH based on CXR alone, with an accuracy of 80%. Internal & external validation is needed before clinical adoption in the clinic.

PD-0731 Improvement of a deep learning based automatic delineation model using anatomical criteria T. Brion 1 , E. Karamouza 2 , L. De Vitry 3 , A. Lombard 3 , T. Roque 3 , N. Paragios 3 , G. Auzac 1 , A. Lamrani-Ghaouti 4 , N. Bonnet 4 , E. Limkin 1 , M. Ung 1 , S. Bockel 1 , D. Pasquier 5 , S. Wong 6 , H. trialists 4 , S. Achkar 1 , S. Rivera 7,1