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

Treatment of cervical cancer commonly involves high-dose-rate brachytherapy (HDR-BT), a procedure that requires precise and efficient planning to achieve the best patient outcomes. Historically, the HDR-BT planning process has been labor-intensive and largely dependent on the expertise of the clinician, resulting in potential inconsistencies in the quality of treatment. To overcome this issue, we propose an innovative method that employs advanced deep-learning models to improve HDR-BT planning. This paper presents the Dendrite Cross-Attention UNet (DCA-UNet), which features a sophisticated dendritic structure comprising a primary branch for stacked inputs and three auxiliary branches dedicated to the segmentation of the clinical target volume (CTV), bladder, and rectum. This architecture enhances the model's understanding of organ-at-risk (OAR) areas, thereby improving dose prediction accuracy. Extensive evaluations reveal that DCA-UNet significantly enhances the precision of HDR-BT dose predictions across different applicator types. Our findings indicate that DCA-UNet consistently outperforms both traditional UNet and the more recent SwimUNetr models. By advancing the use of cross-attention mechanisms in deep learning frameworks, this research aids in the standardization of HDR-BT planning and opens up promising possibilities for future advancements in cervical cancer care.