ESTRO 2025 - Abstract Book

S2425

Physics - Autosegmentation

ESTRO 2025

Results:

Our method achieved convergence in just 14 epochs, reaching a validation accuracy of 43.7%, while the benchmark nnUNet required 1,000 epochs to achieve 50% accuracy. Loop SDE UNet provides both knowledge and data uncertainty estimates: data uncertainty is generated directly by the network, while knowledge uncertainty is calculated by averaging variance across 10 forward passes. Figure 2 shows the epistemic (knowledge) and aleactoric (data) uncertainty map of an input where we can see the high uncertainty area for the networks. By contrast, nnUNet estimates uncertainty using an ensemble method, training 10 models and generating variance for each test case.

Figure 2: A Sample

Conclusion: Our model’s ability to separately estimate data and knowledge uncertainties allows further refinement of the model, enhancing performance and robustness. The explicit separation of uncertainties provides clinicians with detailed uncertainty maps, a valuable tool for prioritizing manual review and improving radiotherapy planning workflows.

Keywords: GTV, uncertainty estimation, segmentation

References: [1]Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany [2] Aerts, H. J. W. L., Wee, L., Rios Velazquez, E., Leijenaar, R. T. H., Parmar, C., Grossmann, P., Carvalho, S., Bussink, J., Monshouwer, R., Haibe-Kains, B., Rietveld, D., Hoebers, F., Rietbergen, M. M., Leemans, C. R., Dekker, A., Quackenbush, J., Gillies, R. J., Lambin, P. (2014). Data From NSCLC-Radiomics (version 4) [Data set]. The Cancer