S1614
Physics - Autosegmentation
ESTRO 2026
imaging for improved GTV p auto-segmentation in head and neck cancer Ruiyan Ni 1 , Levi Burns 2 , Andrew Hope 2,3 , Benjamin Haibe-Kains 1,3 , Alexandra Rink 1,3 1 Department of Medical Biophysics, University of Toronto, Toronto, Canada. 2 Department of Radiation Oncology, University of Toronto, Toronto, Canada. 3 Princess Margaret Cancer Centre, University Health Network, Toronto, Canada Purpose/Objective: Incorporating clinical examination findings into treatment planning is critical for accurate target delineation in several cancer sites. In our previous study, we developed MMINTS (Multi-Modal network with Imaging and Notes for Target Segmentation) and demonstrated its improved performance for auto- segmentation of cervical brachytherapy targets by integrating magnetic resonance (MR) images with clinical notes1. Similarly, the physical and endoscopic examination is essential in head and neck cancers (HNC). Accurate delineation of the Gross Tumor Volume of the primary tumor (GTVp) requires integrating endoscopic examination findings, which capture mucosal and submucosal tumor extent not visible on computed tomography (CT) images2. In this study, we extended MMINTS to a large and heterogeneous HNC dataset with clinical consultation notes to investigate whether incorporating endoscopy information improves GTVp auto-segmentation performance compared to imaging alone. Material/Methods: We used the RADCURE dataset, consisting of CT imaging and GTVp contours for 2969 HNC patients treated with radiation therapy (Figure 1). For each patient, endoscopy findings were extracted from the initial oncology consultation report. These findings were summarized via a large language model (Llama- 3-8B), which was then encoded as numerical vectors using an embedding model (BiomedBERT). The textual embeddings were integrated into the 3D U-Net segmentation network through a cross-attention module applied to the decoder layers, conditioning the image feature decoding on relevant clinical findings. An image-only model was trained with the identical image preprocessing and data splits to serve as a baseline for comparison. Auto-segmentation performance was evaluated on a test cohort of 594 patient GTVp contours using the Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95).
Conclusion: In this study we utilized an custom nnU-Net based model for cardiac substructures delineation. A more detailed analysis of cardiac substructures and their dosimetric parameters, using validated tools for automatic contouring, allows for a more accurate risk assessment and improved prediction of overall survival in this group of patients. References: [1] Bradley JD, Paulus R, Komaki R, Masters G, Blumenschein G, Schild S, et al. Lancet Oncol 2015;16:187–99. https://doi.org/10.1016/S1470- 2045(14)71207-0.[2] Bradley JD, Hu C, Komaki RR, Masters GA, Blumenschein GR, Schild SE, et al. J Clin Oncol 2020;38:706– 14. https://doi.org/10.1200/JCO.19.01162.[3] van der Pol LHG, Blanck O, Grehn M, Blazek T, Knybel L, Balgobind BV, et al. Radiother Oncol 2025;202. https://doi.org/10.1016/j.radonc.2024.11061 0.[4] van der Pol LHG, Pomp J, Hoesein FAAM, Raaymakers BW, Verhoeff JJC, Fast MF. Phys Imaging Radiat Oncol 2024;32. https://doi.org/10.1016/j.phro.2024.100686 Keywords: radiotoxicity, NSCLC, OS prediction
Digital Poster Highlight 4934 Multi-modal integration of endoscopic findings and
Made with FlippingBook - Share PDF online