2639 - Denoising Diffusion Probabilistic Models for Deliverable Radiotherapy Plans

Purpose/Objective(s):

Treatment planning for VMAT remains time- and expertise-intensive, and many AI approaches focus on predicting 3D dose distributions that still require downstream optimization to produce a deliverable plan. Translating dose-level predictions into machine-deliverable control-point parameters remains a key barrier to clinical translation, particularly under modulation and delivery constraints. We propose a conditional diffusion model with a machine-constrained decoder to generate VMAT plan representations that can be directly translated into deliverable control-point parameters.

Materials/Methods:

A total of 386 APBI patients treated in our clinic with VMAT were included. VMAT plan generation was formulated as conditional generation of a resampled full arc fluence representation. Target output was a two-channel 3D tensor corresponding to counterclockwise and clockwise arcs, resampled to 128 × 128 × 128; clockwise arc ordering was reversed to align angular progression between arcs. Inputs comprised CT, multi-channel structure masks; CT and dose were masked to the external-body contour prior to training. A conditional denoising diffusion probabilistic model with a cosine noise schedule (250 steps) was trained to predict the added noise (epsilon prediction) using an L1 loss. The denoiser was a 3D residual U-Net with self-attention (base channels 32; one residual block per resolution level). Predicted fluence was converted to deliverable VMAT control points using a machine-constrained decoder that optimized MLC leaf positions and control-point weights to match fluence while enforcing feasibility via constraints on leaf speed, MU bounds/smoothness (dose-rate feasibility), and aperture geometry regularization (minimum gap and leaf-edge smoothness).

Results:

Training to 50,000 optimization steps converged with stable validation performance, achieving best validation loss of 4.68 × 10^-4 at 31,000 steps. On held-out testing (N=59), predicted fluence showed close agreement with clinical references (MAE 0.011 ± 0.010, median 0.006; SSIM 0.746 ± 0.300, median 0.931). Predicted arc fluence was converted to deliverable control points using a machine-constrained decoder; decoded MLC trajectories satisfied imposed deliverability constraints with smooth inter–control-point motion and no nonphysical leaf behavior, and produced fluence projections consistent with clinical references. Residual discrepancies were concentrated in highly modulated regions where feasibility constraints limited fine-scale modulation.

Conclusion:

A conditional diffusion model with machine-constrained decoder can be structured to generate VMAT plan representations that are both clinically relevant and deliverability-oriented. This approach directly addresses a major translational gap between AI dose prediction and clinically deliverable planning and supports future work toward human-supervised, efficient intelligent treatment planning workflows.