Using deep learning to predict beam-tunable Pareto optimal dose distribution for intensity-modulated radiation...

Area of Science:

• Medical Physics
• Artificial Intelligence
• Radiation Oncology

Background:

• Deep learning models are increasingly used for predicting clinical and Pareto optimal dose distributions.
• Existing models for Pareto optimal dose prediction typically use static beam orientations.
• Predicting Pareto optimal dose distributions for intensity-modulated radiation therapy (IMRT) prostate planning with variable beam numbers and orientations remains an underexplored area.

Purpose of the Study:

• To develop and compare two deep learning models for predicting Pareto optimal dose distributions in IMRT prostate planning.
• To enable real-time prediction of optimal dose distributions using patient anatomy and variable beam configurations.
• To investigate the efficacy of different input modalities for beam angle representation in deep neural networks.

Main Methods:

• Generated 35,000 Pareto optimal IMRT plans for 70 prostate cancer patients using fluence map optimization.
• Developed and compared two deep learning models (Model I and Model II) using anatomical structures (PTV, OARs, body).
• Model I used beam angles as a binary vector input, while Model II converted beam angles into PTV-conformal doses. Models were trained, validated, and tested on patient data, with Mean Square Error (MSE) as the loss function and Adam optimizer.

Main Results:

• Both deep learning models accurately predicted voxel-level dose distributions and dose volume histograms (DVHs) matching ground truth.
• Model I demonstrated significantly lower prediction errors across various metrics (confirmation, homogeneity, R50, D95, D98, D50, D2) compared to Model II.
• Model I also showed superior performance in minimizing mean and maximum dose errors for the planning target volume (PTV) and organs at risk (OARs).

Conclusions:

• The developed deep learning models can accurately predict Pareto optimal dose distributions for IMRT prostate planning.
• Model I, which directly incorporates beam angles, offers improved accuracy and is a promising approach for real-time treatment planning.
• These methods represent a significant advancement towards automated IMRT treatment planning, allowing real-time control over plan optimization trade-offs.