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

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Accurate radiotherapy dose distribution is crucial for head and neck cancer treatment.
• Anatomical variability and critical organ proximity pose challenges in dose prediction.
• Convolutional neural networks (CNNs) are being explored to improve prediction accuracy.

Purpose of the Study:

• To perform a meta-analysis of CNNs for predicting dose distribution in head and neck radiotherapy.
• To evaluate the predictive efficacy of different CNN architectures for planning target volume (PTV) and spinal cord doses.
• To identify factors influencing CNN performance in radiotherapy dose prediction.

Main Methods:

• Systematic literature search of MEDLINE, Embase, Scopus, and Scholar databases.
• Meta-analysis of prediction errors (MAE) for PTV D95 and spinal cord Dmax.
• Assessment of methodological quality using the PROBAST tool.
• Subgroup analyses based on radiotherapy technique, network design, and cancer type.

Main Results:

• Advanced CNNs achieved higher accuracy for PTV D95 prediction (MAE: 0.95), but with significant heterogeneity.
• Classic CNNs provided more consistent spinal cord Dmax predictions (MAE: 0.95) with lower heterogeneity.
• TomoTherapy, specific network designs, and NPC cases showed variations in prediction performance.

Conclusions:

• CNNs show potential for accurate radiotherapy dose prediction, with task-specific customization recommended.
• Advanced CNNs are superior for PTV prediction, while classic CNNs are more consistent for spinal cord prediction.
• Standardized reporting is essential for reproducibility and clinical translation of CNN models in radiotherapy.