High-dose-rate Brachytherapy Planning with Dendrite Cross-Attention UNet

Area of Science:

• Medical Physics
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer treatment often relies on high-dose-rate brachytherapy (HDR-BT).
• Traditional HDR-BT planning is manual, time-consuming, and prone to variability.
• Improving planning efficiency and accuracy is crucial for optimal patient outcomes.

Purpose of the Study:

• To introduce an advanced deep learning model, Dendrite Cross-Attention UNet (DCA-UNet), for automated HDR-BT planning.
• To enhance the precision of dose prediction in cervical cancer brachytherapy.
• To standardize HDR-BT planning and improve treatment consistency.

Main Methods:

• Development of the DCA-UNet architecture with a dendritic structure and cross-attention mechanisms.
• Utilizing auxiliary branches for segmenting clinical target volume (CTV), bladder, and rectum.
• Extensive evaluation of DCA-UNet's performance in HDR-BT dose prediction across various applicator types.

Main Results:

• DCA-UNet demonstrated superior precision in HDR-BT dose predictions compared to traditional UNet and SwimUNetr models.
• The model effectively improved the understanding of organ-at-risk (OAR) areas, leading to more accurate dose predictions.
• Consistent performance enhancements were observed across different applicator types.

Conclusions:

• DCA-UNet offers a significant advancement in automated HDR-BT planning for cervical cancer.
• The proposed deep learning approach contributes to standardizing treatment planning and improving accuracy.
• This research paves the way for future innovations in AI-driven cervical cancer care.