ESTRO 2025 - Abstract Book

S2424

Physics - Autosegmentation

ESTRO 2025

753

Digital Poster Use Stochastic Differential Equations to Disentangle Data and Knowledge Uncertainty in Gross Tumor Volume Contouring Chuxin Zhang 1 , Ana M Barragan Montero 1 , John A Lee 1,2 1 MIRO, UCLouvain, Brussels, Belgium. 2 ICTEAM, UCLouvain, Louvain-la-Neuve, Belgium Purpose/Objective: Accurate Gross Tumor Volume (GTV) segmentation is the starting point of radiotherapy, directly impacting contouring, planning, and delivery steps, and ultimately influencing treatment quality. Clinically, GTV segmentation faces two key challenges: improving efficiency and minimizing variability. To address these, we propose an automated deep learning approach for GTV segmentation that quantifies both data and knowledge uncertainties through Stochastic Differential Equations (SDE). By disentangling these uncertainties, our method allows us to analyze sources of variability, assess segmentation quality, support human adjustments, and further improve model performance. Material/Methods: Our method, Loop SDE UNet (as shown in Figure 1), is based on a classic UNet architecture [1] and SDE-Net [3] with a specialized bottleneck designed to simulate stochastic processes using the Euler-Maruyama method: X t+Δt = X t +f(X t ,t)Δt + g(X t ,t)ΔW t In this formulation, f performs the core segmentation task with deterministic outcomes, while g introduces stochastic elements by modeling data noise. The choice of SDE allows us to separately model deterministic and stochastic elements, with the loop mechanism enabling temporal evolution simulation of the SDE using minimal parameters. This architecture quantifies knowledge uncertainty by measuring prediction variance across multiple passes and directly generates data uncertainty through noise simulation.

Figure 1: Loop SDE UNet Architecture We trained and validated the model on a non-small cell lung cancer (NSCLC) dataset [2] comprising 422 patients, focusing specifically on GTV-1 contours. As a benchmark comparison, we also trained nnUNet on this dataset to evaluate accuracy.