Breast cancer dose distribution prediction based on deep joint learning.

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Healthcare

Background:

• Radiotherapy planning requires balancing dose delivery to the planning target volume (PTV) and protecting organs-at-risk (OARs).
• Accurate dose prediction is crucial for optimizing treatment efficacy and minimizing toxicity in breast cancer radiotherapy.

Purpose of the Study:

• To develop and evaluate a joint learning mechanism for progressive refinement of dose predictions for PTVs and OARs.
• To enhance the accuracy of three-dimensional dose distribution prediction in Intensity-Modulated Radiation Therapy (IMRT) for breast cancer.

Main Methods:

• A novel model constructs a dose prediction network for PTVs, with its output feeding into a subsequent network for overall dose distribution prediction.
• The framework incorporates distance transformation and a dual attention module to boost prediction accuracy.
• The model was trained and validated on a dataset of 307 postoperative breast cancer patients undergoing IMRT.

Main Results:

• The proposed method significantly outperformed the baseline 3DU-Net, showing improvements in dose score, DVH score, ΔPCI, PSNR, SSIM, and various dose metrics (ΔD95, ΔD98, ΔD99, ΔDmax).
• Directly predicting the entire dose distribution map in the second stage proved superior to cascaded networks predicting PTV and OAR maps separately.
• Quantitative improvements included a 0.169 Gy dose score improvement and a 1.382 dB PSNR increase.

Conclusions:

• The developed method generates highly accurate 3D dose distribution maps for breast cancer radiotherapy.
• These accurate dose maps can serve as initializations, thereby enhancing the overall efficiency of radiotherapy planning.