ESTRO 2024 - Abstract Book

S3081

Physics - Autosegmentation

ESTRO 2024

References:

The study is supported by the AIRC grant IG 25951.

[1] Men K. et al. Physica Medica 50 (June 2018), p. 13–19. doi: 10.1016/j.ejmp.2018.05. 006.

[2] Dolezal J. M. et al. Nature Communications 13.1 (Nov. 2022). doi: 10.1038/s41467- 02234025-x.

[3] Wang G. et al.Neurocomputing 338 (Apr.2019), p.34–45.doi: 10.1016/j.neucom.2019.01.103.

[4] Poortmans P. M. et al. The Breast 49 (Feb. 2020), p. 194–200. doi: 10.1016/j.breast.2019.11.011.

[5] Zhong Y. et al. Journal of Applied Clinical Medical Physics (Mar. 2023). doi: 10.1002/ acm2.13951.

[6] Tudda A. et al. Radiotherapy and Oncology 175 (Oct. 2022), p. 10–16. doi: 10.1016/j.radonc.2022.07.012.

[7] Ubeira-Gabellini et al. Manuscript in preparation

1986

Digital Poster

External validation of a deep learning prostate MR auto-contouring model

Maram Alqarni 1,2 , Emma Jones 1,3 , Vinod Mullassery 4 , Stephen Morris 4 , Hema Verma 5 , Sian Cooper 4 , Teresa Guerrero Urbano 4 , Andrew P. King 1 1 King’s College London, School of Biomedical Engineering and Imaging Sciences, London, United Kingdom. 2 Imam Abdurhamn bin Faisal University, Biomedical Engineering, Dammam, Saudi Arabia. 3 Guy’s and St Thomas’ NHS Foundation Trust, Medical physics, London, United Kingdom. 4 Guy’s and St Thomas’ NHS Foundation Trust, Department of Clinical Haematology and Oncology, London, United Kingdom. 5 Guy’s and St Thomas’ NHS Foundation Trust, Radiology, London, United Kingdom

Purpose/Objective:

The success of radiotherapy (RT) treatment for prostate cancer depends largely on the accuracy of prostate contouring for radiation dose planning and optimization. Magnetic resonance imaging (MRI) is an essential aid to volume delineation and the use of MRI based planning is becoming more widespread. Deep learning (DL) models have been proposed to automate this process, but their use is hindered by the problem of domain shift. For example, DL models trained using data from a single scanner vendor/field strength may not generalise well to data from other scanners. The main objective of this work was to validate the performance of a DL autocontouring model trained using heterogeneous public data containing multiple sources of domain shift on an external validation dataset from a UK hospital.