Domain knowledge driven 3D dose prediction using moment-based loss function | JoVE Visualize

Area of Science:

Medical Physics
Radiation Oncology
Artificial Intelligence in Medicine

Background:

Accurate 3D dose prediction is crucial for optimizing intensity modulated radiation therapy (IMRT) plans, especially for lung cancer.
Integrating clinical dose-volume histogram (DVH) constraints into deep learning (DL) models for dose prediction remains challenging due to non-convexity and non-differentiability.
Existing methods like Mean Absolute Error (MAE) loss lack direct incorporation of DVH knowledge, while MAE + DVH loss can be computationally intensive.

Purpose of the Study:

To introduce a novel, convex, and differentiable moment-based loss function for predicting 3D dose distributions in lung IMRT.
To demonstrate that the proposed moment-based loss function can seamlessly integrate DVH domain knowledge into DL frameworks without significant computational overhead.
To compare the performance of the moment-based loss function against traditional MAE and MAE + DVH loss functions.

Main Methods:

A UNet-like convolutional neural network was trained on a dataset of 360 lung cancer patient plans.
The model utilized CT scans, planning target volumes, and organ-at-risk contours as input to predict voxel-wise 3D dose distributions.
Three loss functions were evaluated: MAE, MAE + DVH, and the proposed MAE + moments loss, with performance assessed using DVH metrics, dose-score, and DVH-score.

Main Results:

The MAE + moments loss function significantly improved the DVH-score by 11% compared to MAE loss, with comparable computational cost.
Compared to MAE + DVH loss, the MAE + moments loss achieved an 8% improvement in DVH-score and a 48% reduction in computational cost.
The proposed moment-based approach offers a computationally efficient and mathematically rigorous method for incorporating DVH information.

Conclusions:

The novel moment-based loss function provides a superior and computationally efficient approach for 3D dose prediction in lung IMRT.
This method effectively integrates DVH metrics into DL models, enhancing prediction accuracy and clinical relevance.
The open-source availability of code and models facilitates broader adoption and further research in AI-driven radiation therapy planning.