S2290
Physics - Machine learning and AI algorithms
ESTRO 2026
Purpose/Objective: Inconsistent and institution-specific naming of regions of interest (ROIs) in radiotherapy planning remains a major barrier to automation, multi-center collaboration, and AI reproducibility. Even with the AAPM TG-263 recommendations, clinical datasets often contain hundreds of heterogeneous ROI names due to legacy systems and local conventions. This inconsistency hampers data sharing and limits the generalizability of AI-based planning and dose prediction tools. We aimed to develop and validate a foundation model–driven workflow that performs automated ROI classification and renaming in accordance with TG-263, while allowing flexible customization to institution-specific conventions without model retraining. Material/Methods: A four-stage modular pipeline integrating a large language model (DeepSeek R1) and a vision-language CLIP model was designed. The system sequentially performs:(1) Anatomical site identification,(2) ROI type classification into four categories (Target, OAR, Dose- specific, Auxiliary),(3) Customizable standardization based on TG-263 or local naming protocols, and(4) Laterality verification through CLIP-based image–text alignment.The workflow operates on standard DICOM RTSTRUCT and CT data and requires no retraining. It was evaluated on 600 patients from three institutions and five disease sites (head and neck, breast, lung, cervix, rectum). For each institution, the same protocol was applied directly to assess cross-site generalizability. A separate experiment tested adaptability to customized naming formats (e.g., dose suffixes in Gy, full-word laterality labels). Results: Across all datasets, the system achieved an overall classification accuracy of 99.1% and a renaming accuracy of 98.5%. Performance remained consistent across institutions and disease sites. Minor deviations mainly arose from ambiguous suffixes or missing mappings in custom sheets. The CLIP-based verification module successfully detected and corrected all left–right inversion errors in a stress test involving 36 bilateral OARs. Under customized naming conditions, the workflow maintained 97.7% accuracy, demonstrating strong adaptability to institution- specific formats without retraining or prompt tuning.
Conclusion: This study demonstrates a clinically adaptable and retraining-free foundation model workflow for ROI standardization in radiotherapy. By combining text reasoning with image-based verification, the system achieves high accuracy, robust laterality validation, and flexible customization across institutions. Such a framework enables reproducible AI development, multi-center data harmonization, and seamless integration into automated planning and quality assurance pipelines. Keywords: large language model, ROI renaming Proffered Paper 2120 From Clinical Data to Personalized Risk: A Multimodal AI Workflow for Improving Clinical Decision-making. Davide Dalfovo 1,2 , Carolina Sassorossi 3,4 , Annalisa Campanella 3,4 , Edoardo Mercadante 5 , Filippo Gallina 5 , Luca Boldrini 6,4 , Antonella Martino 6,4 , Maria Serpone 6,4 , Maria Antonietta Gambacorta 6,4 , Alessandra Cancellieri 7,4 , Emilio Bria 8,4 , Rocco Trisolini Trisolini 9,4 , Anna Simbirova 1,2 , Alex Zwanenburg 1,10 , Esther G C Troost 1,10 , Stefano Margaritora 3,4 , Steffen Löck 1,11 , Filippo Lococo 3,4 1 OncoRay, National Center for Radiation Research in Oncology, Faculty of Medicine, Dresden, Germany. 2 Department of Radiotherapy and Radiation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany. 3 Thoracic Surgery Unit, A. Gemelli University Hospital Foundation IRCCS, Rome, Italy. 4 Gemelli University Hospital, Catholic University of the Sacred Heart, Rome, Italy. 5 Division of Thoracic
Made with FlippingBook - Share PDF online