ESTRO 2025 - Abstract Book

S2550

Physics - Autosegmentation

ESTRO 2025

Conclusion: This work showed that DL models are able to detect GTVs with cross-institution recall of 0.67-0.72 with median DSCs of 0.27-0.28. Adding T2w images with ADC enhances cross-institution DSC if detected and recall for center 1, but not for center 2. Consistency of DSC scores over all models and data indicates the MRRN model to be highly generalizable (figure 2).

Keywords: cross-institutional, MRI, deep-learning

References: [1] Simeth J, Jiang J, Nosov A, Wibmer A, Zelefsky M, Tyagi N, Veeraraghavan H. Deep learning ‐ based dominant index lesion segmentation for MR ‐ guided radiation therapy of prostate cancer. Medical Physics. 2023 Aug;50(8):4854-70.

4166

Digital Poster Enhanced Efficiency in GTV Delineation: Evaluating AI 'Active Contouring' Tool

Remus C Stoica 1,2 , Andrei M Dicianu 3 , Andrei Ciobanu 4 , Bogdan Mocanu 5,2 , Mihai Toma 6 , Adrian Radu 7,2 , Bianca M Cotoi 8 , Alexandru Zariosu 7,2 , Maria Timpea 7,2 , Sabina Sucuri 5,2 , Lucian Bicsi 9 , Marius Stãnescu 9 , Dragos Duse 9 , Dragos Grama 9 1 Radiation Oncology, Global Medical Health, Bucharest, Romania. 2 Medical Department, Synaptiq, Cluj-Napoca, Romania. 3 Radiation Oncology, St. Nectarie Oncology and Radiotherapy Center, Craiova, Romania. 4 Radiation Oncology, Amethyst Radiotherapy Center, Timisoara, Romania. 5 Radiation Oncology, Coltea Clinical Hospital, Bucharest, Romania. 6 Radiation Oncology, Amethyst Radiotherapy Center, Otopeni, Romania. 7 Radiation Oncology, Bucharest Institute of Oncology Prof. Dr. Alexandru Trestioreanu, Bucharest, Romania. 8 Physics Department, St. Nectarie Oncology and Radiotherapy Center, Craiova, Romania. 9 Research Department, Synaptiq, Cluj-Napoca, Romania Purpose/Objective: The integration of artificial intelligence (AI) in the automatic contouring of organs-at-risk (OARs) has significantly improved the clinical workflow. Instead of manually delineating OARs, clinicians now correct and refine the AI generated contours. However, this process can be time-consuming, particularly for gross tumor volumes (GTVs) where AI-based auto-contouring solutions are limited due to high variability of tumor locations, shapes, sizes. To