S1615
Physics - Autosegmentation
ESTRO 2026
primary tumour Clinical Target Volumes (CTV-P) in laryngeal, hypopharyngeal, oropharyngeal and oral cavity squamous cell carcinoma: AIRO, CACA, DAHANCA, EORTC, GEORCC, GORTEC, HKNPCSG, HNCIG, IAG-KHT, LPRHHT, NCIC CTG, NCRI, NRG Oncology, PHNS, SBRT, SOMERA, SRO, SSHNO, TROG consensus guidelines. Radiother Oncol. 2018;126(1):3- 24. Keywords: Multi-modal, endoscopic findings, GTV segmentation
Digital Poster 4958
Results: GTVp volumes showed a wide range across testing patients and were categorized into three groups: small (<10 cc), medium (10-30 cc), and large (>30 cc) (Figure 2). Preliminary results demonstrated that the multi- modal MMINTS network achieved improved GTVp auto-segmentation compared to the image-only baseline (overall DSC: 0.71 ± 0.25 vs. 0.68 ± 0.23; HD95: 11.42 ± 10.67 vs. 13.37 ± 11.40). Across the small, medium, and large volume groups, MMINTS yielded mean DSCs of 0.55, 0.73, and 0.75, respectively, compared to 0.54, 0.69, and 0.72 in the baseline model.
Performance Evaluation of a Deep Learning Auto- Segmentation Model for Intracranial Metastases
on Contrast-Enhanced T1-Weighted MRI Norina Predescu 1 , Mirela Dobre 2 , Bianca
Homorozeanu 3 , Gregory Bolard 1 , Szabolcs Botond Lőrincz-Molnár 1 , Francesco Morosato 1 , Anaïs Stefani Stefani 4 , Vincent Marchesi 4 , Jean-Christophe Faivre 4 , Baozhu Sun 5 , Jarkko Niemelä 1 , Jani Pehkonen 1 1 R&D Department, MVision AI, Helsinki, Finland. 2 Radiology, Emergency Clinical Hospital, Bucharest, Romania. 3 Radiation Oncology, Institute of Oncology, Cluj-Napoca, Romania. 4 Radiation Oncology, Institut de Cancérologie de Lorraine Centre Alexis-Vautrin, Department of Radiation Oncology, Vandoeuvre les Nancy, France. 5 Radiation Oncology, Baylor College of Medicine, Houston, USA Purpose/Objective: Due to modern local and systemic therapies, cancer patient outcomes have improved, while the incidence of intracranial metastases has increased.1,2 Stereotactic radiosurgery is a validated treatment for brain metastases.3,4Precise delineation of intracranial metastases is critical for radiotherapy treatment planning and dose delivery. Manual segmentation can be time-consuming and prone to inter-observer variation. This study presents an internal evaluation of a deep learning (DL) auto- segmentation model designed to assist in radiotherapy workflows by automatically detecting and segmenting intracranial metastases on contrast- enhanced T1-weighted MRI scans. Material/Methods: A test dataset was collected from two partner clinics in the US and France. A total of 148 MRI scans from 148 adult patients were retrospectively included in the test set. Each scan was annotated by a trained medical professional with medical annotation experience and underwent two independent peer-review rounds, by a Radiation Oncologist and a Radiologist, to ensure annotation consistency and accuracy. The final dataset contained 1,002 manually delineated metastases, serving as ground truth. The DL model was evaluated
Conclusion: We applied the multi-modal MMINTS network to incorporate endoscopy findings into GTVp auto- segmentation for HNC. MMINTS achieved higher geometric accuracy compared to the baseline. This study enables the model to access complementary information typically available only to human experts, demonstrating the potential of multi-modal learning to improve clinical reliability in radiotherapy across disease sites. References: 1. Ni R, Haibe-Kain B, Rink A. Integrating vaginal involvement information from clinical notes into deep learning-based CTVHR segmentation for cervical HDR brachytherapy. Brachytherapy. 2025;24(4):S50.2. Grégoire V, Evans M, Le QT, et al. Delineation of the
Made with FlippingBook - Share PDF online