Multi-centre generalisability of deep learning-based dose prediction for head and neck radiotherapy using DAHANCA...

Area of Science:

• Radiotherapy
• Medical Physics
• Oncology

Background:

• Deep learning shows promise for automating radiotherapy planning and quality assurance in head and neck cancer (HNC).
• Clinical implementation requires robust validation of model generalisability across diverse datasets.
• This study assesses a deep learning dose prediction model's generalisability in a multi-centre HNC cohort.

Purpose of the Study:

• To evaluate the generalisability of a deep learning dose prediction model trained on single-centre data.
• To compare the model's performance against a median-based prediction model.
• To explore the impact of dose prediction accuracy on expected patient toxicity.

Main Methods:

• A deep neural network was trained on 388 HNC treatment plans.
• The model was evaluated on an internal test set (42 patients) and a national DAHANCA cohort (560 plans).
• Performance was assessed using dose metrics (Dmean) and compared to a median model, with toxicity explored via NTCP and NTI.

Main Results:

• The deep learning model closely matched clinical dose distributions, with small median Dmean differences for PTVs and OARs across cohorts.
• The model significantly outperformed the median-based model, which exhibited larger dose variations.
• Predicted toxicity showed modest differences from clinically planned toxicity, suggesting potential for quality assurance insights.

Conclusions:

• The deep learning dose prediction model demonstrated strong generalisability across multiple institutions, outperforming a median-based approach.
• Prediction-plan differences were within observed interquartile ranges, supporting clinical adoption.
• Discrepancies in predicted toxicity may highlight areas for improving clinical radiotherapy plans and quality assurance.