Deep-learning Method for the Prediction of Three-Dimensional Dose Distribution for Left Breast Cancer Conformal...

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Increasing demand for advanced radiotherapy planning systems.
• Learning-based approaches offer potential for improved planning speed and quality.
• Deep learning models present a novel perspective for radiotherapy planning.

Purpose of the Study:

• To evaluate the accuracy and precision of a deep learning model (U-Res-Net) for predicting 3D dose distribution in breast cancer radiotherapy.
• To compare the model's predictions against a standard treatment planning system.

Main Methods:

• A 3D U-Res-Net model was trained and tested using CT images and contouring data from 120 breast cancer patients.
• The model's output (predicted dose distribution) was compared to dose distributions calculated by a treatment planning system for 10 test patients.
• Key metrics included Dice similarity coefficients, dose difference, Mean Absolute Error (MAE), and gamma passing rates.

Main Results:

• The deep learning model achieved an average Dice similarity coefficient of 0.91 ± 0.03 for isodose volumes.
• Average MAE for test cases was 5.71 ± 1.19%, with an average dose difference of 0.60 ± 2.81% across all voxels.
• 3D gamma passing rates with 3 mm/3% criteria ranged from 78.99% to 97.58% for planning target volume and organs at risk.

Conclusions:

• Deep learning models can accurately predict 3D dose distributions for left breast cancer patients.
• The developed model demonstrates optimal accuracy and precision in radiotherapy planning.
• Future research can extend this model to predict dose distributions for other cancer types.