ESTRO 2025 - Abstract Book

S3403

Physics - Machine learning models and clinical applications

ESTRO 2025

3. de Biase A, Sourlos N, van Ooijen PMA. Standardization of Artificial Intelligence Development in Radiotherapy. Semin Radiat Oncol. 2022;32:415 – 20. 4. Modelcard Creator - a Hugging Face Space by AI4MIRO [Internet]. [cited 2024 Nov 18]. Available from: https://huggingface.co/spaces/AI4MIRO/Model_Cards_Writing_Tool

2608

Proffered Paper Development of a probabilistic model for metastatic lymphatic progression in early-stage breast cancer Zineb Smine, Alex de Caluwé, Antoine Desmet, Nick Reynaert, Jennifer Dhont Medical Physics, Institut Jules Bordet, Brussels, Belgium Purpose/Objective: Randomized trials have shown mixed results on which lymph node levels (LNLs) to include in CTVs for elective radiotherapy of early-stage breast cancer (eBC). In this study, we propose a personalized approach to CTV definition using a probabilistic Hidden Markov Model (HMM) for lymphatic tumor progression in axillary LNLs 1-3 depending on primary T-category and observed macroscopic metastases, extending prior work in head and neck cancer to a new pathology [1]. Material/Methods: A HMM graph tailored to eBC was constructed where the five directed arcs reflect lymphatic flow from the tumor to each LNL, and between LNLs (Figure 1). To calculate the probabilities of tumor spread along each of the five arcs, pathology data from Rosen et al. [2], derived from post-mastectomy lymph node dissections, was used. The dataset provided individual LNL involvement in early and late T-category BC in 339 and 594 patients, respectively. As in the model by Ludwig et al. [1] the presence/absence of microscopic metastasis in each LNL was modeled via one hidden binary random variable (Figure 1). The hidden state is linked to an observable binary variable indicating the presence/absence of macroscopic metastases on any imaging modality via its sensitivity and specificity. We illustrate the use of the produced HMM for risk assessment of metastases in LNLs 1-3 in a breast cancer patient for whom PET-CT imaging is used for lymph node staging (specificity of 90% and sensitivity of 64%) [3].

Results: