1266 - Dual-Model Deep Learning Framework for Accurate Benchmark Dose Prediction In Brain Stereotactic Radiosurgery

Purpose/Objective(s): Dose gradient in brain SRS directly impacts normal tissue toxicity, yet no standardized benchmark currently exists for cross-platform plan quality assessment. We developed a size-stratified deep learning framework to predict patient-specific benchmark dose, employing appropriately scaled volumes of interest (VOI) to ensure accurate dose gradient prediction across the full range of lesion sizes.

Materials/Methods: Benchmark Gamma Knife (GK) plans optimized for maximal dose conformality were generated for 461 patients (1,097 targets). Our deep learning framework consisting of two 3D U-Net models with attention mechanisms was trained: a small target model (for volume <1 cc, 756 targets, median 0.27 cc [range 0.002-1.0], 64×64×64 mm³ VOI, 5-fold cross-validation ensemble) and a large target model (for volume =1 cc, 341 targets, median 2.9 cc [range 1.0-75.3], 128×128×128 mm³ VOI). A custom loss function prioritized accuracy in the dose falloff region, i.e., the 1-cm ring around the target volume. Performance was evaluated on an independent test set of 137 patients with 209 targets (103 small, 106 large) and compared against two single-model baselines using paired Wilcoxon signed-rank tests.

Results: On 209 independent external test targets, the small target model (volume < 1 cc, n=103) achieved a mean absolute error (MAE) of 0.43±0.15 Gy in the dose falloff region, V_50% Dice of 0.866, and |V_12Gy error| of 0.09±0.13 cc. The large target model (volume = 1 cc, n=106) achieved an MAE of 0.35±0.17 Gy in the dose falloff region, V_50% Dice of 0.909, and |V_12Gy error| of 0.72±1.57 cc. Compared to single-model baselines, the dual-model approach achieved 30% lower intra-target MAE and significantly better Dice, and V_12Gy accuracy overall (all p<0.001). The large VOI baseline model performed 26% worse in falloff MAE (p<0.001), while the small VOI baseline model showed equivalent falloff accuracy but significantly worse intra-target prediction.

Conclusion: The dual-model deep learning framework accurately predicted patient-specific benchmark dose distributions for brain SRS. It achieved statistically significant improvements relative to single-model alternatives, enabling reliable cross-platform plan-quality evaluation and direct utility for routine clinical SRS quality assurance for all lesion sizes.