The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
technical-relevance esophageal deep-learning treatment-planning

Deep learning dose prediction met clinical constraints in esophageal radiotherapy planning

A conditional generative adversarial network predicted clinically acceptable esophageal radiotherapy dose distributions, but systematic target underdosage limited clinical use.

RadOncToday · ·1 min read

KEY POINTS

  • Feasibility study of a conditional generative adversarial network-based model for three-dimensional dose prediction in esophageal cancer intensity modulated radiotherapy.
  • The dataset included 110 esophageal cancer patients; 95 plans were used for training and 15 for testing.
  • All predicted plans met clinical dose constraints, giving a 100% qualification rate.
  • Mean prediction error across dose-volume histogram parameters was 2.88 ± 2.88%, with lower target error of 1.00 ± 1.00%.
  • Predicted plans had comparable homogeneity index (0.08 ± 0.02 versus 0.08 ± 0.01) but lower conformity index than manual plans (p = 0.0034); mean Dice similarity coefficient across isodose levels was 0.916.

CLINICAL TAKEAWAY

This model may be useful as a preliminary planning-reference tool for esophageal intensity modulated radiotherapy, especially for suggesting achievable dose-volume histogram constraints and identifying dose-limiting organs at risk. It is not ready for autonomous clinical plan generation: only 6 of 15 predicted plans achieved the highest target-dose grade, target underdosage was systematic, and the test set was small and single-institutional.

SOURCE

Journal of Applied Clinical Medical Physics