Flexible-Cm GAN: Towards Precise 3D Dose Prediction in Radiotherapy

Riqiang Gao, Bin Lou, Zhoubing Xu, Dorin Comaniciu, Ali KamenCVPR 2023 · DOI

Introduction

In modern radiotherapy (RT), the primary objective is to deliver a lethal radiation dose to a tumor while minimizing exposure to surrounding healthy Organs at Risk (OARs). This process, known as Knowledge-Based Planning (KBP), is traditionally iterative and time-consuming. While deep learning has advanced automation in this field, existing models often lack the flexibility required for real-world clinical use. They typically assume fixed beam geometries or treatment modes, making them "brittle" when clinicians need to adjust parameters like beam angles or switch between static Intensity-Modulated Radiation Therapy (IMRT) and dynamic Volumetric Modulated Arc Therapy (VMAT).

Comparison of 2D image translation and 3D dose prediction across different beam modes.

The Flexible-C$^m$ GAN (FCGAN) addresses these gaps by providing a conditional generative model that handles varying treatment configurations and incorporates clinical constraints directly into the learning process.

The Flexible-C$^m$ GAN Framework

FCGAN treats 3D dose prediction as a multi-conditional image-to-image translation task. Unlike standard models, it accepts a variety of inputs, including CT scans, PTV/OAR masks, and spatial "plates" representing beam angles and planning modes.

The FCGAN architecture featuring multi-condition inputs and specialized loss functions.

Key Methodological Innovations

  • Miss-Consistency Loss ($L_{mc}$): Clinical datasets are frequently incomplete. FCGAN uses a miss-consistency loss to ensure that the model generates consistent dose maps regardless of whether specific beam information is provided or imputed. This allows the model to leverage heterogeneous datasets effectively.
  • Shift-Dose-Volume (SDV) Loss ($L_{sdv}$): The Dose-Volume Histogram (DVH) is the clinical gold standard for plan evaluation but is inherently non-differentiable. The authors derived a differentiable SDV loss that translates frequency-level DVH errors into voxel-level gradients, allowing the model to prioritize high-dose regions critical for tumor control.
  • Interactive Inference: The conditional nature of the GAN allows for real-time updates. Clinicians can adjust beam angles through a graphical interface and observe the predicted 3D dose distribution almost instantaneously.

Clinical Evaluation and Results

The model was validated using two lung cancer datasets: D1WA (365 patients with beam data) and D2NA (793 patients without beam data). A "toy" evaluation on the FFHQ face dataset also demonstrated the model's ability to handle missing attributes like glasses or gender by treating them as conditional inputs similar to beam parameters.

In clinical testing, FCGAN significantly outperformed state-of-the-art baselines such as DeepDoseNet and DoseGAN. On the D1WA dataset, it achieved a 7.3% reduction in Shift-DV Error (SDE). Qualitatively, the predicted dose maps showed superior conformity to the target tumor volume compared to existing automated methods, closely mimicking plans generated by expert human dosimetrists.

Conclusion and Future Directions

FCGAN represents a significant step toward practical, automated radiotherapy planning. By bridging the gap between computer vision techniques and medical physics constraints, it offers a flexible and robust tool for clinicians. Future work will focus on expanding validation to other anatomical sites, such as the prostate and head-and-neck, and moving toward end-to-end planning by predicting machine-executable fluence maps directly from the dose distributions.

Summary updated

BibTeX
@inproceedings{gao2023flexible,
  author = {Riqiang Gao and Bin Lou and Zhoubing Xu and Dorin Comaniciu and Ali Kamen},
  title = {{Flexible-Cm GAN: Towards Precise 3D Dose Prediction in Radiotherapy}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  doi = {10.1109/cvpr52729.2023.00076},
}
Figure