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

This is the authors' abstract. We don't add key points for this paper.

Introduction Deep learning dose prediction shows promise for automated radiotherapy planning and quality assurance in head and neck cancer (HNC). Clinical adoption requires validation of model generalisability using clinically relevant metrics. The study evaluated generalisability of a dose prediction model trained on single-centre data in a large multi-centre cohort. Materials and methods A deep neural network was trained on 388 HNC treatment plans from one Danish centre and evaluated on an internal test dataset of 42 patients and on a national DAHANCA cohort of 560 plans from six institutions. Performance was evaluated using dose metrics and compared with a median model assigning each OAR the median D mean from the training data. Expected toxicity was explored using NTCP and the normalised toxicity index (NTI). Results Predictions closely matched clinically applied dose distributions with median D mean differences of −0.1 Gy [interquartile range −0.5, 0.3] for PTVs and 1.1 Gy [-0.6, 3.7] for OARs across cohorts. The median-based model showed larger interquartile ranges, with D mean differences of 0.7 Gy [0.0, 2.1] for PTVs and −4.6 Gy [-14.6, 3.9] for OARs. Comparison of expected toxicity between model-predicted and clinically applied dose distributions showed median NTI differences of 4.4% [ 0.4, 9.4] and median NTCP differences of about 1–3% for xerostomia and dysphagia grade 2 + . Conclusions The dose prediction model outperformed the median-based model, with prediction-plan differences within interquartile ranges across cohorts, supporting generalisability. Discrepancies between model-predicted and clinically planned toxicity may indicate suboptimality in clinical plans and could inform quality assurance.