ESTRO 2025 - Abstract Book

S248

Brachytherapy - Gynaecology

ESTRO 2025

Results: Based on the best trade-off between reconstruction accuracy and time, the '3D_lowres' nnU-net configuration was selected. On the 36 test plans, the average DSC was 0.94 ± 0.04 (0.76 - 0.97), and an average DPE was 0.76 ± 0.52 (0.30 - 3.52) (Fig. 2). Differences between the Venezia and Geneva applicators were small (Fig. 2). Average automated reconstruction time was 51.3 ± 6.0 sec (37.5 sec for segmentation and 13.8 sec for registration), with no significant differences between Venezia and Geneva applicators.

Conclusion: A deep learning-based workflow for automated applicator reconstruction in MRI-based cervical cancer brachytherapy has been developed, achieving on average a 0.76 mm agreement between automatically reconstructed dwell positions and clinical, ground-truth positions. With an average reconstruction time of 51 seconds, automated reconstruction substantially reduces overall fraction planning times and labor demands, while reducing risk of applicator movement.