Standardization of breast radiotherapy structure nomenclature using transformer-based multi-modal feature learning

Résumé fourni par la source

Standardized nomenclature for radiotherapy (RT) organs-at-risk (OARs) and target volumes (TVs) is essential for using large retrospective imaging and treatment-planning datasets. However, clinical structure names vary substantially across centres, making manual standardization time-consuming and resource-intensive. This study aims to develop a transformer-based multimodal model to automatically standardize the breast RT structure names into protocol-standard labels by combining raw clinical-use text, image features, transformer-derived spatial context, and geometric features. The model was trained and internally validated on a curated dataset from a single centre comprising 1436 breast cancer RT patients, and externally tested on 463 patients from five independent centres. During internal validation, the full multimodal model achieved a weighted F1-score of 98.08 ± 0.09%. Across the five external centres, the full model achieved weighted F1-scores ranging from 87.75% to 95.40%, with OAR accuracy reaching 100% in most centres and 96% in one centre. For primary and nodal TVs, external accuracy ranged from 80% to 97%. With reduced training data, the transformer-based model outperformed the CNN baseline when trained using 50%–80% of the available training data, with the largest difference observed at 50% training data (88.12% vs 75.36% weighted F1). These results support transformer-based multimodal learning as a scalable approach for retrospective RT nomenclature standardization, with potential application in clinical quality assurance workflows, offering a substantial reduction in manual effort and time.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.