Adaptive Voxel Importance Prediction for Rapid Radiotherapy Treatment Planning

Abstract

Large-scale inverse optimisation remains a practical bottleneck in online adaptive radiotherapy, where treatment replanning must be completed within minutes. Existing voxel-reduction methods can accelerate optimisation, but many rely on patient-specific probing or hand-crafted heuristics before each plan. This work presents AdaVIP, an extension of the voxel-importance framework of Mair et al. that replaces iterative probing with one-pass voxel-importance prediction. AdaVIP combines frozen nnU-Net anatomical features with compact physics-informed descriptors to rank voxels for reduced optimisation. Stochastic inverse-probability weighting preserves the target objective in expectation for the sampling stage, while post-sampling safeguards improve coverage of clinically critical regions. Across prostate, liver, and head-and-neck planning cohorts, AdaVIP reduces planning time by about sevenfold relative to full optimisation while keeping major dose–volume histogram endpoints within predefined clinical equivalence margins. The method adds less than one second of inference time and preserves compatibility with the downstream convex optimisation workflow.