Deep learning-assisted needle artifact suppression for enhanced anatomical visualization in prostate high-dose-rate brachytherapy ultrasound imaging.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiation Oncology, Stanford University, Palo Alto, CA. Electronic address: [email protected].
- Department of Radiation Oncology, Stanford University, Palo Alto, CA.
Abstract
Implanted needles introduce acoustic artifacts that degrade transrectal ultrasound (TRUS) images during high-dose-rate (HDR) prostate brachytherapy, complicating ultrasound-only contouring. We developed an artificial intelligence (AI) needle eraser to remove needles and associated artifacts and facilitate structure delineation. TRUS images from 120 patients undergoing HDR prostate brachytherapy were retrospectively collected. Three-dimensional volumes were acquired immediately before and after needle insertion. A Cycle Generative Adversarial Network (CycleGAN) was trained to transform postneedle images into needle-free images. Two physicians independently rated clinical utility for prostate and urethra delineation using a four-point scale (1 = poor; 4 = excellent). Prostate and urethra contours from preneedle and needle-erased images were compared with clinical reference contours using Dice similarity coefficient (DSC). The AI tool suppressed needles and associated artifacts, improving visualization of the prostate and urethral lumen. Mean reader scores were 3.80 ± 0.43 for prostate and 3.75 ± 0.37 for urethra delineation. Compared with preneedle contours, needle-erased contours showed higher agreement with clinical references: prostate DSC increased from 0.91 to 0.93, and urethra DSC increased from 0.70 to 0.89. A CycleGAN-based needle eraser can generate needle-free ultrasound images from postinsertion scans and improve visualization and contour agreement for physician delineation.