Autosegmentation Via nnU-Net on Post-Brachytherapy-Catheter CT for Prostate HDR Brachytherapy
Abstract
Purpose
To develop and validate a 3D deep learning autosegmentation model for prostate and OAR contouring on post-brachytherapy-catheter CT images for HDR prostate brachytherapy planning.
Methods
A self-configuring U-Net architecture (nnU-Net) was trained on 206 HDR prostate brachytherapy cases. An independent test set of 36 patients was used for evaluation. Structures included the prostate PTV, bladder, rectum, and urethra. Model performance was compared against a commercially available solution originally trained for external beam radiotherapy (EBRT), using clinically approved, physician-drawn contours as ground truth. Evaluation metrics included geometric metrics such as Dice coefficient (DSC), Hausdorff distance (HD95%), and average surface distance (ASD). Clinically relevant dosimetric measures, including PTV V100%, bladder and rectum D1cc, and urethra D0.1cc, were also assessed. Statistical significance was determined using paired Wilcoxon signed-rank tests.
Results
The nnU-Net model demonstrated superior geometric accuracy for all structures within slices containing the prostate. For prostate PTV, nnU-Net achieved DSC = 0.91 +/- 0.04, HD95% = 3.62±2.23 mm, and ASD = 1.25±0.66 mm, compared to DSC = 0.78±0.07, HD95% = 9.52±4.73 mm, and ASD = 2.97±0.90 mm for the commercial model. Urethra segmentation was only available with nnU-Net (DSC = 0.84±0.09). Segmentations produced by nnU-Net resulted in clinically negligible dosimetric changes, including ΔPTV V100% = -0.25±2.30%, bladder ΔD1cc = 0.13±0.75 Gy, rectum ΔD1cc = 0.34±0.74 Gy, and urethra ΔD0.1cc = -0.03±0.16 Gy. No significant difference in PTV V100% was observed for nnU-Net (p=0.91), while the commercial model showed significant deviations (p < 0.001).
Conclusion
A brachytherapy-specific nnU-Net model enables accurate and reliable autosegmentation of the prostate and OARs on post-brachytherapy-catheter CT images. Compared to an EBRT-trained commercial solution, it offers improved geometric performance with minimal dosimetric impact, supporting its clinical feasibility and potential to streamline HDR prostate brachytherapy workflows.