3659 - Efficient Robustness Optimization in Intensity Modulated Proton Therapy for Head and Neck Cancer via Dual Attention Gated Visual State Space Generative Adversarial Networks (DAGVSS-GAN)

Purpose/Objective(s): In contrast to intensity-modulated photon therapy (IMRT), intensity-modulated proton therapy (IMPT) plans do not depend on the planning target volume (PTV). Instead, they directly delineate tumor volumes by utilizing the clinical target volume (CTV) and gross tumor volume (GTV). Robust optimization in IMPT must consider patient positioning errors and uncertainties in the proton ion range. The positioning uncertainties for patients with head and neck cancers are 0.3 cm (isotropic), and the ion range uncertainties stand at 3.5%. To address these uncertainties, assessing 21 different dose distribution scenarios is essential. Even when high-performance servers are employed, this evaluation process is extremely time-consuming. If more substantial errors are present, an even greater number of scenarios would need to be considered. Implementing a deep learning method is highly advisable to boost planning efficiency, minimize patient waiting times, and streamline the treatment workflow.

Materials/Methods: We developed the DAGVSS-GAN, which integrates a visual state space module into the network generator. We also incorporated spatial attention and channel attention blocks into the discriminator. Additionally, we designed a novel "NT Gamma" loss function that separately accounts for dose deposition location errors and numerical discrepancies. For the experimental dataset, we used data from 157 brain tumor patients who had undergone proton therapy. The dataset was divided into a training set of 117 patients and a validation set of 40 patients.

Results: Quantitative evaluations and statistical analyses confirmed a high level of consistency between the generated dose distributions and the reference doses from the treatment planning system (TPS). For the single-beam dose, the 3D overall gamma passing rates (GPRs) were 98.67%±1.12% (3mm/3%, 10% threshold) and 97.80%±1.31% under stricter criteria (2mm/2%, 10% threshold). For the entire-plan dose, the 3D overall gamma passing rates reached 98.49%±1.17% (3mm/3%, 10% threshold) and 97.33%±1.23% under stricter criteria (2mm/2%, 10% threshold). The average differences in target coverage (TC), dose selectivity (DS), gradient index (GI), and homogeneity index (HI) were 0.31±0.15%, 0.04±0.02, 0.09±0.01, and 0.07±0.02, respectively.

Conclusion: The results show that the DAGVSS-GAN effectively optimizes IMPT dose distributions, achieving high consistency with the treatment planning system's outcomes. It improves the speed of robust-proton-dose optimization, boosts physicists' efficiency, and streamlines the proton therapy workflow.