Quality-Driven Deep Learning Dose Prediction for Head and Neck Cancer Using Limited High-Quality Training Data
Abstract
Purpose
To evaluate whether deep learning models trained on a small number of high-quality plans (e.g., ≤30) can predict dose distributions of comparable quality, and whether the predicted quality improvements are achievable.
Methods
This study includes 138 oropharyngeal VMAT patients, whose clinical treatment plans were scored using a validated percentile-based plan quality scoring system developed in-house (PQS; higher scores indicate superior plan quality). Three training strategies were compared: (1) the baseline model trained with 84 cases spanning all quality levels, (2) the model trained with only 30 top-scoring cases (PQS≥65), and (3) the model trained with only 30 top-scoring cases after transfer learning from 110 external automated-planning cases to address limited training data. The baseline model was evaluated on its 26-patient test set; the high-quality models were applied to the 88 patients outside the high-quality subset. Two cases underwent dosimetrist replanning guided by predicted dose distributions and structure-specific quality metrics derived from them.
Results
The baseline and high-quality-only strategies predicted negative PQS changes (−6.64±16.95 and −9.46±18.84 percentile points), reflecting regression toward average quality. The high-quality with transfer learning strategy predicted positive changes (+8.20±21.34 points; p<0.001), with 67% of patients showing predicted improvement. In a paired comparison of 16 overlapping test patients, the transfer learning approach produced higher predicted PQS in 14/16 cases (87.5%), with a mean advantage of 9.38±8.87 percentile points (p<0.001). Dosimetrist replanning confirmed predictions were achievable, improving PQS from 7.97 to 81.2 and from 44.20 to 70.3.
Conclusion
Despite training on only 30 high-quality cases, models using transfer learning predicted achievable quality improvements rather than reproducing average practice, providing actionable optimization targets for clinical planners.