Demo: Generative AI helps Radiotherapy Planning with User Preference

TL;DR

This work presents the Flexible Dose Proposer (FDP), a two-stage, slider-conditioned generative model that enables interactive, region-specific 3D radiotherapy dose prediction. Stage I stabilizes learning with a VQ-VAE foundation, while Stage II conditions on CT, structures, and user-preference sliders encoded via adaptive normalization to produce deliverable doses aligned with HI, CI, PTV, and OAR goals. Integrated with the Eclipse TPS, FDP translates 3D dose predictions into optimization objectives and demonstrates improved adaptability and plan quality over Varian RapidPlan in head-and-neck cases, including real-time UI demonstrations. While validated primarily on head-and-neck data, the approach illustrates a practical pathway toward preference-driven AI-assisted radiotherapy planning within clinical workflows.

Abstract

Radiotherapy planning is a highly complex process that often varies significantly across institutions and individual planners. Most existing deep learning approaches for 3D dose prediction rely on reference plans as ground truth during training, which can inadvertently bias models toward specific planning styles or institutional preferences. In this study, we introduce a novel generative model that predicts 3D dose distributions based solely on user-defined preference flavors. These customizable preferences enable planners to prioritize specific trade-offs between organs-at-risk (OARs) and planning target volumes (PTVs), offering greater flexibility and personalization. Designed for seamless integration with clinical treatment planning systems, our approach assists users in generating high-quality plans efficiently. Comparative evaluations demonstrate that our method can surpasses the Varian RapidPlan model in both adaptability and plan quality in some scenarios.

Figures (8)

• Figure 1: Our flexible dose proposer pipeline. Stage I pre-trains a foundational dose decoder, regularizing the dose prediction to be realistic. The second stage trains a flexible dose prediction model with heterogeneous conditions.
• Figure 2: Estimated vs. achieved DVHs of RapidPlan and our FDP model. RapidPlan estimations are not well aligned with achieved plans, may due to its limited generalization to unseen patients.
• Figure 3: Demonstration of user preference with sliding bars. The left/right of sliding bars represents better/worse quality for the specific structure respectively. The Preference 1 (P1) has the preferences of OAR sparing over PTV homogenity, and Preference 2 (P2) has the opposite. The preferences have been captured in the dose prediction, as shown in the DVHs in the right panel.
• Figure 4: Example demonstrates the value of Stage I pre-training.
• Figure 5: Screenshot of the Gradio demo interface. A full demo video recording can be found in an https://huggingface.co/HappySubmit/DoseProposerDemo/tree/main.