A multi-toxicity deep learning approach for normal tissue complication probability modelling in head and neck cancer patients receiving radiotherapy

PURPOSE: Toxicities after radiotherapy for head and neck cancer (HNC) often co-occur and share underlying mechanisms, yet most conventional and deep learning (DL) NTCP models predict only a single endpoint. By developing DL NTCP models which can predict multiple toxicities simultaneously, this study aimed to capture inter-toxicity relationships to improve prediction performance.

METHODS AND MATERIALS: A multi-institutional cohort of 1,418 HNC patients was used to develop and validate a multi-toxicity (MT) DL model, incorporating 3D dose distributions, CT scans, organ-at-risk segmentations and patient-related features, that simultaneously predicts five toxicities; aspiration, dysphagia, sticky saliva, taste alteration and xerostomia, all evaluated six months after treatment. Results are compared to conventional NTCP models, as well as a set of single-toxicity (ST) 3D DL models.

RESULTS: The MT model outperformed both the conventional and ST models for dysphagia (AUC = 0.83 versus 0.81 and 0.82) and xerostomia (0.80 versus 0.75 and 0.78) prediction on the independent validation cohort. The latter models achieved better performance for sticky saliva (0.72 and 0.71 versus 0.69) and taste alteration (both 0.67 versus 0.71). The MT model achieved a higher AUC on aspiration than the ST model (0.71), but performed as well as the reference model (both 0.74). Within the external validation cohort, all models performed comparably to each other, with the MT model achieving a slightly higher average AUC (0.64) across all endpoints than the conventional and ST models (both 0.63). Sub-analyses revealed that the benefit of the proposed multi-toxicity modelling varied by endpoint.

CONCLUSION: MT models offer comparable-and in some cases improved-performance over conventional single-endpoint approaches, indicating their promise for NTCP modelling. However, benefits are not uniform across all endpoints, highlighting the importance of considering toxicity-specific features when designing multi-toxicity models.