Dendrite cross attention for high-dose-rate brachytherapy distribution planning

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Oncology

Background:

• Cervical cancer treatment relies heavily on high-dose-rate brachytherapy (HDR-BT).
• Manual HDR-BT plan creation is time-consuming and lacks standardization due to planner expertise variability.
• Accurate segmentation of clinical target volume (CTV) and organs at risk (OAR) is critical for effective HDR-BT planning.

Purpose of the Study:

• To introduce and evaluate two novel deep learning models, BiCA-UNet and DCA-UNet, for automated HDR-BT dose prediction.
• To enhance the accuracy and standardization of HDR-BT treatment planning through advanced AI.
• To improve the correlation between CT scans, segmentations, and dose predictions in HDR-BT.

Main Methods:

• Development of Bi-branch Cross-Attention UNet (BiCA-UNet) integrating CT scans and segmentations for improved CTV analysis.
• Introduction of Dendrite Cross-Attention UNet (DCA-UNet) with a dendritic structure for enhanced OAR segmentation and dose prediction refinement.
• Utilizing cross-attention mechanisms within both models to improve feature representation of critical anatomical structures.

Main Results:

• Both BiCA-UNet and DCA-UNet demonstrated significant improvements in HDR-BT dose prediction accuracy across various applicator types.
• The cross-attention mechanisms effectively enhanced feature representation of CTV and OARs, leading to more precise dose predictions.
• The models showed potential for standardizing HDR-BT planning and improving reliability.

Conclusions:

• BiCA-UNet and DCA-UNet represent advanced deep learning solutions for HDR-BT planning.
• These models can enhance treatment plan accuracy, standardization, and potentially improve patient outcomes in cervical cancer.
• Further research into these AI-driven approaches holds promise for the future of radiation oncology.