Deep learning auto-contouring of target and organs-at-risk on post-catheter implant CT images for prostate HDR brachytherapy.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiation Medicine, University of Wisconsin-Madison, Madison, Wisconsin, USA.
- Department of Medical Physics, University of Wisconsin-Madison, Madison, Wisconsin, USA.
Abstract
Accurate delineation of the prostate and surrounding organs-at-risk (OARs) is essential for HDR prostate brachytherapy. Manual contouring on post-catheter CT images is time-consuming and prone to variability due to artifacts and anatomical deformation from the implanted brachytherapy catheters. To develop and validate a 3D deep learning autosegmentation model for prostate and OAR contouring on post-catheter CT images for prostate brachytherapy planning. A self-configuring U-Net architecture (nnU-Net) was trained on 206 HDR prostate brachytherapy cases and tested on an additional 36 patients. Structures included prostate PTV, bladder, rectum, and urethra identified from a Foley catheter. Performance was evaluated against a commercially available model using geometric (Dice coefficient (DSC), Hausdorff distance (HD95%), average surface distance (ASD)) and dosimetric metrics (PTV V100%, bladder and rectum D1cc, urethra D0.1cc). Statistical significance was assessed using paired Wilcoxon signed-rank tests. NnU-Net achieved superior geometric accuracy versus the commercial model for all structures within slices containing the prostate. For prostate PTV, nnU-Net yielded 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 provided by nnU-Net (DSC = 0.84 +/- 0.09). Dosimetric differences for nnU-Net were clinically negligible: 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. Differences in PTV V100% were not significant for nnU-Net (p = 0.91) but significant for the commercial model (p < 0.001). A brachytherapy-specific nnU-Net model enables accurate autosegmentation of prostate and OARs on post-catheter CT images, outperforming an EBRT-trained commercial solution. Minimal dosimetric differences support clinical feasibility and potential workflow improvements in HDR prostate brachytherapy.