Catheter position prediction using deep-learning-based multi-atlas registration for high-dose rate prostate...

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Medicine

Background:

• Catheter placement in high-dose-rate (HDR) prostate brachytherapy is currently empirical and physician-dependent.
• This variability can lead to uncertainty in treatment plan quality and suboptimal dose distribution.

Purpose of the Study:

• To develop and evaluate a learning-based method for guiding catheter placement in HDR prostate brachytherapy.
• To reduce physician dependence on experience and improve prostate cancer treatment plan quality.

Main Methods:

• A framework combining deformable registration (Reg-Net), multi-atlas ranking, and catheter regression was developed.
• Distance maps of prostate and organs-at-risk were used for registration and atlas ranking based on similarity and catheter position criteria.
• A retrospective study on 90 patients with fivefold cross-validation evaluated the method's feasibility and impact on dose metrics.

Main Results:

• The proposed method achieved comparable dose coverage (V100 = 95%) to clinical plans.
• Plans from predicted catheter patterns showed slightly increased hotspots (V150 by 5.0%, V200 by 2.9%) on average.
• Organ-at-risk dose constraints showed minimal average differences (within ±1 cc) for bladder, rectum, and urethra.

Conclusions:

• A novel deep-learning-based multi-atlas registration algorithm for catheter placement prediction in HDR prostate brachytherapy was developed.
• The method shows significant clinical potential for improving treatment plan quality and consistency.
• Further clinical evaluation is warranted to validate its role in quality control for HDR prostate brachytherapy.