Radiation exposure and clinical validation of autosegmentation models for the supraventricular cardiac conduction system in breast cancer radiotherapy: an institutional perspective

Background: Radiation dose to cardiac conduction nodes may contribute to arrhythmia risks in breast cancer (BC) patients after radiotherapy, yet dosimetric evidence remains limited. This study aimed to evaluate doses to the sinoatrial (SAN) and atrioventricular nodes (AVN) in BC patients treated with intensity-modulated radiation therapy (IMRT) and to clinically validate a deep learning-based autosegmentation model for these structures. Methods: A retrospective analysis was conducted on 87 BC patients who underwent IMRT. Doses to the whole heart, four cardiac chambers, the SAN, and the AVN were evaluated and correlated. For autosegmentation, a convolutional neural network (CNN) was trained on 60 patients, validated on seven, and tested on 20. Segmentation accuracy was assessed using the Dice similarity coefficient (DSC), and dosimetric consistency was compared between automated and manual contours. Results: for SAN: 0.63 in left- and right-sided cases; for AVN: 0.77 and 0.63, respectively). The autosegmentation model achieved DSCs of 0.83 for SAN and 0.75 for AVN, with no statistically significant dosimetric differences between autosegmented and manual contours. Conclusions: The SAN receives substantial irradiation in right-sided BC patients during IMRT, and RA dose strongly correlates with conduction node doses, suggesting its potential as a clinical surrogate. The CNN-based autosegmentation method enables accurate and efficient delineation of the SAN and AVN, facilitating reliable dosimetric assessment in clinical practice.