ESTRO 2024 - Abstract Book

S2994

Physics - Autosegmentation

ESTRO 2024

[4] Buda, Mateusz. “U-Net for brain segmentation”, GitHub repository, https://github.com/mateuszbuda/brain segmentation-pytorch

[5] Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., & Schmidt, L. “OpenCLIP”, Computer software, https://doi.org/10.5281/zenodo.5143773

176

Digital Poster

Automated CTV delineation for prostate salvage radiotherapy using deep learning

Daniel Rusche 1 , Marco M E Vogel 1 , Johannes Kiechle 1,2,3 , Lucas Etzel 1 , Stephanie E Combs 1,4,5 , Jan C Peeken 1,4,5

1 Klinikum rechts der Isar, Technical University of Munich (TUM), Department of Radiation Oncology, Munich, Germany. 2 Technical University of Munich, Institute for Computational Imaging and AI in Medicine, Munich, Germany. 3 Munich Data Science Institute, Konrad Zuse School of Excellence in Reliable AI, Munich, Germany. 4 Deutsches Konsortium für Translationale Krebsforschung, Partner Site Munich, Munich, Germany. 5 Institute of Radiation Medicine, Helmholtz Zentrum, epartment of Radiation Sciences, Munich, Germany

Purpose/Objective:

Clinical target volume (CTV) delineation poses a time-consuming task for radiation oncologists. Though the exact segmentation is essential, there remains a considerable interrater variability. To overcome these obstacles, we aimed to develop an automated CTV for prostate salvage radiotherapy delineation network based on deep learning using the latest ESTRO-ACROP CTV guideline (1).

Material/Methods:

We randomly collected planning computed tomography (CT) scans for salvage prostate radiotherapy from a cohort of 60 patients from the local database at the Department of Radiation Oncology at the Klinikum rechts der Isar. Patients with an initial pathological tumor status of pT3b or higher, rectal balloon, contrast enhancement, endoprosthesis, or scans in abdominal position were excluded. Scans were acquired with the two CT scanners Somatom go.Open Pro or Emotion 16 (Siemens, Germany) with fixed protocols each. The CTV was delineated manually for each case by a resident in accordance with the current ESTRO-ACROP guideline using Varian Eclipse (Varian, USA). After review by an expert (J. P.) the cases were exported and converted with Slicer (version 5.4.0). For automatic segmentation, we trained an nnUNet v2 for each dataset split into 5 folds for cross-validation with the 2D and 3D-fullres setting respectively. The training was performed on a NVIDIA RTX A6000 GPU for 100 epochs. After final validation, the best model was determined along with the possibility of ensembling. Testing performance was evaluated using 12 patients of the cohort which were not included in the training or validation.