ESTRO 2024 - Abstract Book

S3165

Physics - Autosegmentation

ESTRO 2024

1b

Figure 1: Variables importance per classification Models 1 (1a) and 2 (1b), considering the outcome of prediction as either binary (with Dice coefficient≥0.8 indicating a good performance) or continuous.

Conclusion:

Overall, the U-net yielded a satisfactory performance across the dataset. However, accuracy as assessed per the Dice score is decreased in presence of some qualitative radiological features (e.g. irregular margins, complex shape). Further training and external validation on complex lesions may further improve model performance and open doors to clinical implementation.

Keywords: deep learning, nsclc, cnn

3178

Digital Poster

Development of a CT-only autosegmentation model for hippocampal delineation in radiation therapy.

Konrad Stawiski 1,2 , Mateusz Pajdziński 2,3 , Adam Zięba 2 , Michał Masłowski 2 , Jacek Fijuth 3,2

1 Medical Univerasity of Lodz, Department of Biostatistics and Translational Medicine, Lodz, Poland. 2 Copernicus Memorial Hospital, Department of Radiation Oncology, Lodz, Poland. 3 Medical Univerasity of Lodz, Department of Radiotherapy, Lodz, Poland

Purpose/Objective:

Hippocampal avoidance in whole-brain radiation therapy (HA-WBRT) and prophylactic cranial irradiation (HA-PCI) has become the standard of care to mitigate cognitive impairment associated with brain irradiation. However, accurate delineation of the hippocampus typically necessitates high-resolution MRI, which may be challenging to acquire in