1196 - Simultaneous Source Position and Continuous Shield Orientation Optimization for Directional Brachytherapy via Neural Implicit Dose Kernels

05:35pm - 05:40pm ET

Room 160

Presenter(s)

Joseph Schulz, MS, BS - University of Wisconsin Madison, Madison, WI

J. B. Schulz¹, A. Trestrail¹, J. M. Floberg², B. Bednarz¹, and J. Slagowski²; ¹Department of Medical Physics, University of Wisconsin-Madison, Madison, WI, ²Department of Human Oncology, University of Wisconsin-Madison, Madison, WI

Purpose/Objective(s): Directional brachytherapy with shielded sources (IMBT) reduces OAR doses by preferentially attenuating radiation away from critical structures. Current planning discretizes shield orientation into 8-36 fixed angles with combinatorial optimization, reducing plan quality. This study presents the first continuous shield orientation optimization method, enabled by an implicit neural representation (INR) framework, and hypothesizes that joint dwell time, position and continuous orientation optimization achieves superior OAR sparing versus discrete-angle IMBT planning.

Materials/Methods: An INR mapping 3D spatial coordinates to scalar dose was trained on a Monte Carlo water dose kernel from a prototype shielded I-125 source (hemispherical tungsten shield, EGSnrc). Shield angle is parameterized continuously through coordinate rotation of query coordinates into the kernel's canonical frame, making dose fully differentiable with respect to shield orientation. The framework was evaluated retrospectively on 10 prostate HDR patients (mean PTV 36.8cc, range 25.9-54.6; mean 16.6 catheters, range 14-21) treated with Ir-192 at one institution (IRB-approved). Two methods were compared: i) discrete IMBT with directional I-125 at 24 angles, and ii) continuous INR-optimized directional I-125. All plans were normalized to PTV V100=95% with identical catheter count and optimization settings, differing only in angular parameterization. DVH endpoints were PTV V150%<65%, urethra V118%<0.1cc, rectum and bladder V75%<1cc. Wilcoxon signed-rank tests assessed pairwise differences.

Results: The INR dose kernel achieved 0.2% mean and 1.4% max error versus Monte Carlo. Continuous optimization significantly reduced PTV V150% (61.2±11.4 vs 64.1±9.6%, p=0.03) and urethral V118% (0.17±0.23 vs 0.24±0.25cc, p=0.01) compared to discrete IMBT, with improvements in 7/10 and 8/10 patients respectively. Rectal V75% (0.70±0.41 vs 0.71±0.39cc, p=1.00) and bladder V75% (1.15±0.89 vs 1.27±0.93cc, p=0.38) were equivalent. Shield orientations spanned -180° to 177° with per-patient mean 3.0±1.1° and standard deviation of 89.6±4.2°, confirming the optimizer exploits the full continuous angular space.

Conclusion: Continuous orientation optimization via INR dose kernels eliminates the angle quantization bottleneck, establishing a new paradigm for differentiable geometric optimization in brachytherapy planning. Because the framework operates directly on Monte Carlo kernels, it generalizes to any source geometry without TG-43 parameterization, enabling rapid adoption of novel directional source designs. These findings motivate directional source hardware development for prostate brachytherapy and demonstrate a framework for rapid intraoperative plan optimization, while highlighting dosimetric challenges of low-energy emitters for single-isotope HDR brachytherapy. Future work will evaluate application to a directional Yb-169 source.