Abstract
Radiotherapy planning requires balancing lethal doses to tumors with the preservation of healthy organs at risk (OARs). Traditional deep learning models often lack flexibility and suffer from institutional bias. We present the Flexible Dose Proposer (FDP), a two-stage generative framework that enables real-time, interactive dose prediction based on user-defined preferences. By utilizing a foundational dose decoder and a preference-aware encoder, FDP allows clinicians to explore the Pareto front of treatment plans, outperforming existing industry standards in both accuracy and personalization.
Introduction
Radiotherapy planning is a high-stakes balancing act. Clinicians must deliver a lethal dose of radiation to a tumor (the Planning Target Volume, or PTV) while minimizing damage to surrounding healthy tissue (Organs at Risk, or OARs). Currently, deep learning models for dose prediction are often "black boxes" trained on specific institutional datasets, leading to institutional bias and a lack of flexibility. Every patient is unique, and clinicians often need to make subjective trade-offs—for instance, accepting slightly less tumor uniformity to better protect a patient's salivary glands.
The Flexible Dose Proposer (FDP) bridges this gap by offering a range of dose "flavors" that users can adjust interactively, shifting the role of AI from a static predictor to a collaborative tool.
Methods
The FDP framework employs a two-stage generative approach to ensure physical realism and precise user control.
Stage I: Foundational Dose Decoder
Using a Vector Quantized Variational Autoencoder (VQ-VAE), the model is pre-trained on a massive dataset of 31,000 dose distributions. This stage acts as a structural anchor, learning the "language" of radiation physics and ensuring that the final predictions are physically possible and free from boundary artifacts.
Stage II: Flexible Dose Prediction
This stage utilizes a MedNext-style encoder that processes CT images, anatomical contours, and beam configurations. Crucially, it accepts user preference prompts via adaptive instance normalization (AdaIN). An objective loss function ($L_{obj}$) penalizes the model if the predicted dose deviates from the user's requested trade-off level.
Datasets and Evaluation
The model was pre-trained on 31,000 diverse dose distributions to establish physical constraints. For fine-tuning and evaluation, the researchers used six cohorts of head-and-neck cancer cases, totaling over 800 patients. Head-and-neck cases are particularly challenging due to the dense concentration of critical organs near the target volume.
Key Results and Insights
The FDP model demonstrates significant improvements over industry-standard tools like Varian’s RapidPlan™. Unlike 1D Dose-Volume Histogram (DVH) predictors, FDP predicts full 3D dose grids, allowing for a deeper understanding of spatial relationships between organs and tumors.
- Superior Performance: FDP achieved better sparing for 14 out of 15 organs at risk compared to baselines.
- Real-Time Interaction: The model's forward pass takes only 30ms, with full 3D visualization updating in 1.5 seconds, enabling a seamless interactive experience.
- Stability: Ablation studies confirmed that the Stage I pre-training is essential for eliminating unrealistic "boundary artifacts" and maintaining stability when user sliders are pushed to extremes.
Conclusion and Future Work
While currently validated for head-and-neck cancer, the FDP framework represents a paradigm shift toward interactive, personalized radiotherapy. Future work will focus on extending the model to other anatomical sites like the prostate or lung, which present different motion and physical constraints. Clinical adoption will require multi-institutional blind studies to further validate the subjective quality and practical utility of these AI-generated plans in diverse clinical settings.
Summary updated