AI needle removal improves urethral visualization during prostate HDR brachytherapy planning

CycleGAN-generated needle-free ultrasound improved urethral contour agreement from 0.70 to 0.89 Dice after prostate HDR catheter insertion.

KEY POINTS

  • This retrospective single-institution study included 120 patients undergoing ultrasound-guided HDR prostate brachytherapy. Investigators developed a CycleGAN-based deep-learning tool to remove implanted needles and associated artifacts from postinsertion transrectal ultrasound images while preserving anatomy for prostate and urethral delineation.
  • The model was trained using 6,249 ultrasound slices from 100 patients, with 1,225 slices from 20 independent patients reserved for testing. Unpaired image translation allowed the network to learn needle-free image appearance without requiring anatomically matched pre- and postinsertion ultrasound volumes.
  • The algorithm suppressed needle-associated reverberation, comet-tail artifacts, and acoustic shadowing. Two experienced brachytherapy physicians rated its usefulness at 3.80 ± 0.43 out of 4 for prostate contouring and 3.75 ± 0.37 for urethral contouring, although agreement between readers was poorer for urethral assessment.
  • Contours generated using AI-enhanced postimplant images agreed more closely with clinical reference contours than contours transferred from preneedle images. Prostate Dice similarity coefficient increased from 0.91 to 0.93, while mean Hausdorff distance decreased from 5.41 to 4.36 mm.
  • The improvement was more pronounced for the urethra: mean Dice coefficient increased from 0.70 to 0.89, Hausdorff distance decreased from 3.76 to 1.50 mm, and mean distance to agreement decreased from 0.91 to 0.30 mm. These differences highlight the difficulty of transferring contours from preneedle imaging after implantation-induced deformation.
  • Inference required approximately 10.8 seconds per patient on the held-out cohort, making near-real-time use during treatment planning technically feasible. The original postimplant ultrasound remains necessary for catheter reconstruction and implant verification; the enhanced images are intended as an additional contouring reference.
  • Occasional slice-to-slice urethral discontinuities were observed because the model processes two-dimensional images without explicit three-dimensional continuity constraints. There was no independent postimplant MRI/CT ground truth for validating potentially hallucinated anatomy, and no evaluation demonstrated improvements in treatment-plan dosimetry or clinical outcomes.

CLINICAL TAKEAWAY

AI-based artifact suppression may address an important limitation of ultrasound-only HDR prostate brachytherapy by improving urethral visualization after catheter placement without requiring patient repositioning or additional imaging. However, generated anatomy must remain subject to physician verification because geometric discontinuities and image synthesis uncertainty have not been fully resolved, and improvements in contour agreement have not yet been linked to better dosimetry.

SOURCE

Brachytherapy

Browse more research Suggest a correction