Probabilistic dose prediction using mixture density networks for automated radiation therapy treatment planning

Viktor Nilsson, Hanna Gruselius, Tianfang Zhang, Geert De Kerf, Michaël ClaessensPhysics in Medicine & Biology 2021 · OpenReview

TL;DR

This paper introduces Mixture Density Networks (MDNs) for automated radiation therapy treatment planning, enabling probabilistic dose predictions that capture the inherent trade-offs in clinical decision-making. By modeling dose distributions as a mixture of possibilities rather than a single value, MDNs can better represent complex clinical scenarios and support the creation of more consistent and potentially higher-quality treatment plans.

What Problem Does This Paper Address?

Radiation therapy treatment planning (RTTP) is a complex and time-consuming process. It involves balancing competing clinical objectives, such as maximizing radiation dose to the tumor while minimizing it to surrounding healthy organs (organs at risk). This balancing act is often subjective, leading to variability in treatment plans even for similar patients. Existing automated planning methods, while aiming to reduce this variability, often predict a single "optimal" dose distribution. This deterministic approach struggles to represent the inherent uncertainty and the existence of multiple viable treatment strategies that arise from conflicting clinical priorities. The paper addresses this gap by proposing a method that can predict a distribution of possible doses for each voxel, reflecting the range of acceptable outcomes and the trade-offs involved.

Methods

The core of this work is the application of Mixture Density Networks (MDNs). Unlike standard neural networks that output a single value (e.g., a dose for each voxel), MDNs are designed to output the parameters of a probability distribution. Specifically, they use a Gaussian mixture model, which can represent multimodal distributions. This allows the network to learn and predict that a voxel might have different likely dose values, corresponding to different clinical priorities.

The authors employ a U-net architecture, a type of convolutional neural network well-suited for image segmentation and prediction tasks. The input to the U-net consists of binary encodings of relevant anatomical regions (ROIs) and their Euclidean distance transforms. The network is trained to predict the parameters (mean, variance, and mixing coefficients) of a two-component Gaussian mixture model for each voxel.

To demonstrate the practical utility of these probabilistic predictions, a dose mimicking procedure is introduced. This method uses the predicted probability distributions to generate deliverable treatment plans that align with the different "modes" or components of the predicted dose distribution. This allows for the creation of plans that explicitly reflect the different clinical trade-offs learned by the MDN.

Datasets

The study utilizes a dataset of postoperative prostate cancer patients from the Iridium Cancer Network (Antwerp, Belgium). For each patient, two distinct treatment plans were generated:

  • Plan P₁: Prioritizes target coverage.
  • Plan P₂: Prioritizes sparing the rectum.

These dual plans were created using an ML-based dose prediction and mimicking system, resulting in 32 plans in total for the training dataset (16 patients, each with P₁ and P₂). The data was split into training, validation, and holdout sets with a ratio of 50/25/25.

Key Insights and Ideas

  • Probabilistic Prediction Captures Clinical Trade-offs: The most significant insight is that MDNs can effectively learn and represent the inherent uncertainty and multiple viable solutions arising from conflicting clinical priorities (e.g., target coverage vs. organ-at-risk sparing). This moves beyond deterministic predictions to a more realistic representation of clinical decision-making.
  • MDNs as a Decision Support Tool: By providing a distribution of possible doses, MDNs can offer clinicians a comprehensive overview of viable plan alternatives, potentially saving time and facilitating more informed decisions during the RTTP process.
  • Dose Mimicking for Deliverable Plans: The proposed dose mimicking method demonstrates a practical application of probabilistic predictions, showing how to translate these distributions into actual, deliverable treatment plans that reflect different clinical strategies.
  • Separation of Modes Reflects Clinical Intent: Visualizations clearly show that the predicted modes (μ₁ and μ₂) learned by the MDN correspond well to the distinct clinical intents of the P₁ and P₂ ground truth plans, particularly in critical regions like the rectum.

Key Results

  • The MDN successfully learned to predict two distinct modes (μ₁ and μ₂) that closely followed the ground truth dose distributions (P₁ and P₂) for prostate cancer patients, particularly in the critical area where the prostate and rectum overlap.
  • The predicted standard deviations (σ₁ and σ₂) provided a measure of uncertainty within each mode.
  • The dose mimicking procedure generated deliverable plans that spatially followed the predicted modes well, demonstrating the practical utility of the probabilistic approach.
  • Qualitative comparisons of dose distributions and DVHs showed that the predicted modes and mimicked plans captured the essential differences between the high-coverage (P₁) and rectum-sparing (P₂) protocols.

Limitations and Future Work

  • Overfitting: The authors note that the MDN trained for 100 epochs, but incipient overfitting was observed around 50 epochs, suggesting a need for careful hyperparameter tuning or regularization.
  • Conformality of Mimicked Plans: The mimicked plans were slightly less conformal than the predicted distributions, likely due to the predictions being "overly optimistic" about how quickly dose falls off around the target.
  • External ROI Deviation: The dose to the seminal vesicles was somewhat overestimated, leading to a slight deviation in the predicted DVH for the external ROI.
  • Future Work:
    • Investigating MDNs for patients with different treatment protocols or more complex tradeoffs (e.g., different boost volumes, varying NTCP objectives).
  • Exploring MDNs with more than two mixture components for scenarios with more than two distinct clinical priorities.
  • Developing three-dimensional MDNs instead of slice-by-slice prediction.
  • Improving methods to counteract overfitting and stabilize training for high-dimensional problems.
  • Integrating MDN knowledge into multi-criteria optimization frameworks.

Summary updated

BibTeX
@inproceedings{nilsson2021probabilistic,
  author = {Viktor Nilsson and Hanna Gruselius and Tianfang Zhang and Geert De Kerf and Michaël Claessens},
  title = {{Probabilistic dose prediction using mixture density networks for automated radiation therapy treatment planning}},
  booktitle = {Physics in Medicine \& Biology},
  year = {2021},
  url = {https://openreview.net/forum?id=DXOPZmR1a8},
}
Figure