A GPU-accelerated brachytherapy planning framework for template-based applications: Evaluated on three-dimensional...

3D-printed individualized templates improve high-dose-rate (HDR) brachytherapy, but planning remains limited by slow optimization and geometry-driven needle design. This study proposes a GPU-accelerated framework with a fast dose engine and adaptive inverse optimizer to enhance efficiency and plan quality in template-based workflows.

The PyTorch-implemented framework was validated against a commercial TPS via single-source tests and full-plan recalculations. Retrospective re-optimization of 13 cervical cancer patients treated with 3D-printed templates evaluated HR-CTV conformity, coverage, and organ-at-risk (OAR) parameters.

The GPU dose engine showed good agreement with TPS, with differences below 6% in controlled single-source tests and below 1% in full-plan patient dose comparisons. In patient evaluations, the framework significantly improved HR-CTV conformity (+12.5%) and D₁₀₀ values, while maintaining comparable HR-CTV D₉₀ coverage. Bladder high-dose indices (D_0.1cc, D_1cc, D_2cc) were reduced by ∼2-3% (p<0.05). Rectum doses showed no significant differences; sigmoid and small bowel parameters slightly increased (+1-2%, p<0.05) but remained below prescription limits. Optimization required 12-34 seconds per plan.

The GPU-accelerated framework provides accurate, rapid planning with improved HR-CTV conformity and bladder sparing, maintaining clinically acceptable OAR doses. Its design enables integration into 3D-printed template workflows, supporting efficient and personalized cervical cancer HDR brachytherapy.