Deep learning for head and neck radiation dose prediction: a systematic review and meta-analysis

Accurate dose distribution prediction in head and neck radiotherapy remains complex due to anatomical variability and proximity of critical organs. This study presents a meta-analysis of convolutional neural networks (CNNs) and related architectures to evaluate their predictive efficacy. The MEDLINE (PubMed), Embase (Elsevier and Ovid), Scopus, and Scholar databases were searched using a structured search string. EndNote software was used for screening. Meta-analysis was conducted using the effect size of the absolute prediction error (MAE) for D95 in the planning target volume (PTV) and Dmax in the spinal cord (as a representative organ at risk, OAR). The PROBAST tool was used to assess methodological quality. Missing training parameters were retrieved from supplementary sources, and updated analyses were performed. For PTV (D95), advanced CNNs achieved the highest pooled accuracy (MAE: 0.95, 95% CI: 0.62-1.27), though substantial heterogeneity (I² > 99%) limits interpretability. For spinal cord (Dmax), classic CNNs yielded the most consistent predictions (MAE: 0.95, 95% CI: 0.72-1.18) with lower heterogeneity. Subgroup analyses revealed performance variation by radiotherapy technique (e.g., TomoTherapy showed the lowest PTV error), network design (dense block inclusion increased OAR error), and cancer type (NPC cases had the lowest PTV error). Generalization beyond the spinal cord to other OARs remains limited. Advanced CNNs demonstrated superior accuracy in PTV prediction, while classic CNNs performed more consistently for spinal cord estimation. These findings support task-specific model customization based on cancer type, modality, and OAR anatomy, with standardized reporting to enhance reproducibility and clinical translation.