Foundation Model-Driven Regions of Interest Classification and Renaming in Cancer Radiotherapy: A Customizable, Retraining-Free Workflow Across Institutions.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China.
- Department of Oncology, Shanghai Medical College, Fudan University, Shanghai, China.
- Department of Oncology, Xiangya Hospital, Central South University, Changsha, Hunan, China.
- Department of Radiation Oncology, The First Affiliated Hospital of Zhejiang University School of Medicine, Hangzhou, China.
- Radio Therapy Business Unit, Shanghai United Imaging Healthcare Co, Ltd, Shanghai, China.
Abstract
Inconsistent and institution-specific naming of regions of interest (ROIs) in cancer radiotherapy planning limits interoperability, hinders automation, and introduces ambiguity in clinical workflows. This study aims to develop and evaluate a multistage, foundation model-based workflow for automated ROI classification and standardization in accordance with AAPM Task Group 263 (TG-263) nomenclature, while maintaining flexibility for customization to institution-specific naming conventions. A four-stage modular pipeline integrating large language models (LLMs) and a Contrastive Language-Image Pretraining (CLIP) model was developed. The system sequentially performs (1) anatomic site classification, (2) ROI type classification into four categories, (3) standardization based on customizable protocols, and (4) laterality verification via CLIP. The pipeline operates on standard DICOM RTSTRUCT and CT data and requires no model retraining. We evaluated the workflow on 600 patients from three institutions across five disease sites (head and neck, breast, lung, cervix, rectum). Classification and renaming accuracy were assessed at both global and institution-site levels, with error taxonomy used to characterize failure modes. The workflow achieved an overall ROI type classification accuracy of 99.12% and a renaming accuracy of 97.92%. Classification performance was consistently high across all institutions and sites; renaming accuracy showed minor variation driven by protocol omissions, compound suffix ambiguity, and institution-specific naming conventions. CLIP-based laterality verification corrected all errors in a targeted stress test with deliberate left-right label inversions. This study demonstrates that foundation language models, augmented by a vision-language module for laterality verification, can robustly automate ROI classification and naming across institutions, offering a customizable and retraining-free solution for standardizing cancer radiotherapy workflows.