A feasibility study on deep learning-based individualized 3D dose distribution prediction.

Area of Science:

• Medical Physics
• Radiotherapy
• Deep Learning

Background:

• Radiation therapy treatment planning is complex and time-consuming.
• Current methods often rely on trial-and-error.
• Deep learning (DL) models can predict dose distributions but lack personalization for physician preferences.

Purpose of the Study:

• To develop a DL model for predicting individualized 3D radiation therapy dose distributions.
• To incorporate patient anatomy and physician-defined trade-offs (via dose volume histograms - DVH) into dose prediction.
• To enable fine-tuning of dose distributions based on specific clinical goals.

Main Methods:

• A modified U-Net deep learning network was developed.
• Inputs included patient planning target volume (PTV)/organ at risk (OAR) masks and desired DVH.
• The model was trained on data from 77 prostate cancer patients.

Main Results:

• The model accurately predicted Pareto optimal 3D dose distributions.
• Predicted DVHs closely matched the desired DVHs.
• Maximum dose errors were within 3.6% for mean dose and 2.0% for maximum dose.

Conclusions:

• A 3D U-Net model was successfully developed to predict individualized, Pareto optimal dose distributions.
• The model uses patient anatomy and desired DVH curves as inputs.
• Predicted dose distributions can serve as valuable references for clinical treatment planning.