ESTRO 2021 Abstract Book

S1422

ESTRO 2021

performed by 5 expert observers using the median of these observers as gold standard.

Results The results of the internal 5-fold cross validation of the segmentation CNN are shown in Fig. 2 (first column). Clinician-standard delineation is achieved in the parotid glands, spinal cord and mandible. Even though the performance for the brainstem was marginally lower, the segmentation was still very good, with a median HD95 of (3.37 ± 1.50) mm and median mDTA of (0.95 ± 0.37) mm. The segmentation model generalises well in the external cohort of patients and the distance metric measures are shown in Fig. 2 (last column). The CNN achieves similar accuracy as the 5 human observers relative to their median and performs consistently with the internal validation results.

Conclusion Using an automated pre-processing method to crop input data to an anatomically consistent sub-volume enables training of accurate auto-segmentation models with limited data. In our case, a model trained using 34 cases performed within observer variation. The ability to train high accuracy models with small datasets enables protocol-specific models, improving consistency for applications such as clinical trials or retrospective modelling studies. 1. doi: 10.7937/k9/tcia.2020.a8sh-7363 2. https://arxiv.org/abs/1809.04430 3. doi: 10.1186/1748-717X-7-32

Digital Poster: Implementation of new technology, techniques, clinical protocols or trials (including QA & audit)

PO-1696 Clinical commissioning a synchrotron-based proton therapy system J.D. Azcona 1 , B. Aguilar 1 , L. Irazola 2 , A. Viñals 1 , P. Cabello 1 , D. Zucca 1 , A. Perales 1 , R. Polo 1 , J.M. Delgado 1 , J. Burguete 3 1 Clínica Universidad de Navarra, Radiation Physics and Radiation Protection, Madrid, Spain; 2 Clínica Universidad de Navarra, Radiation Physics and Radiation Protection, Pamplona, Spain; 3 School of Sciences,