Current status and future directions of multimodal deep learning normal tissue complication probability modelling in head and neck cancer radiotherapy

Abstract

Normal tissue complication probability (NTCP) modelling is central to balancing tumour control and toxicity risk in radiotherapy (RT) for head and neck cancer (HNC). Despite major advances in radiation delivery and imaging, toxicity remains a major determinant of long-term quality of life in survivors. Classical NTCP models based on dose–volume histograms (DVHs) and phenomenological equations such as the Lyman–Kutcher–Burman model have provided clinically valuable guidance for decades but oversimplify the complex, multidimensional nature of the normal tissue radiation response. Recent developments in machine learning (ML) and deep learning (DL) have enabled the integration of large-scale clinical, imaging, dosimetric and biological data to generate more accurate, patient-specific toxicity predictions. However, data scarcity, endpoint variability, and limited external generalisability have restricted clinical translation. This mini-review summarises the current status of NTCP modelling in HNC, focusing on evolving multimodal DL approaches and data fusion strategies. A structured literature review approach identified 10 contemporary studies of multimodal DL models for radiotherapy-related toxicity prediction in HNC. Across included studies, fusion architectures were inconsistently reported and model development procedures were also variably described regarding feature selection, hyperparameter tuning, model selection and validation. Current multimodal DL NTCP models are promising but not yet ready for routine clinical use, with robust external validation, calibration assessment, reproducibility, clinical utility evaluation and prospective testing required before integration into RT decision-support workflows.