ESTRO 2024 - Abstract Book

S3037

Physics - Autosegmentation

ESTRO 2024

Ronneberger, O., Fischer, P., Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab, N., Hornegger, J., Wells, W., Frangi, A. (eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science(), vol 9351. Springer, Cham. https://doi.org/10.1007/978-3-319-24574-4_28 Roy, A. G., Conjeti, S., Navab, N., & Wachinger, C. (2019). Bayesian QuickNAT: Model uncertainty in deep whole-brain segmentation for structure-wise quality control. In NeuroImage (Vol. 195, pp. 11–22). Elsevier BV. https://doi.org/10.1016/j.neuroimage.2019.03.042

1125

Digital Poster

Variance of auto segmentation of organs at risk in brain cancer patients

Anne H Andresen 1,2 , Laura Toussaint 3,2 , Yasmin Lassen 3 , Slavka Lucakova 1,2 , Christian R Hansen 1,4,5 , Jesper F Kallehauge 1,2 1 Aarhus University hospital, Danish Centre for Particle Therapy, Aarhus, Denmark. 2 Aarhus University, Department of Clinical Medicine, Aarhus, Denmark. 3 Aarhus university hospital, Danish Centre for Particle Therapy, Aarhus, Denmark. 4 University of Southern Denmark, Laboratory of Radiation Physics, Odense, Denmark. 5 Odense University hospital, Department of Oncology, Odense, Denmark

Purpose/Objective:

Auto-segmentation has the potential to play an important role in radiotherapy of patients with brain cancer[YAL1] [1]; however effective strategies to identify and mitigate unreliable segmentations are required for clinical implementation. For organs at risk (OARs) in the brain there have been varying accuracies of deep learning models for multi-organ segmentation due to large differences in OAR sizes [2], [3]. This could reduce the reliability of the model in a clinical setting. Therefore, evaluating the ensemble variance in a model, could serve as a crucial metric for assessing the reliability and generalizability of the model, with a higher variance signifying increased diversity and potential unreliability, and lower variance indicating greater consistency and better generalizability. [4] The aim of this study was twofold 1) we investigated a methodology to generate high-quality segmentations independent of organ size, 2) we investigated the ensemble variance of the final segmentations of the OARs in the brain to assess the variability and reliability of this approach.

Material/Methods:

Two distinct deep learning models were individually trained for the segmentation of eight OARs: nnUNet—a convolutional neural network which uses convolutions on the entirety of the scans—and a transformer model, which gives patches of the scan as sequential inputs to a multi head attention mechanism. The output segmentations of the two models were subsequently combined to generate a final segmentation which was used for evaluation.