Robustness of Transformer-Based Fluence Map Prediction Under Clinically Realistic Perturbations
For the medical physics community, this work highlights the fragility of deep learning-based IMRT planning under distribution shifts and the inadequacy of SSIM for capturing clinically relevant errors.
This paper evaluates the robustness of transformer-based fluence map prediction for IMRT under clinically realistic perturbations, finding smooth degradation under moderate shifts but sharp failures under severe rotations and noise, with hierarchical transformers showing improved robustness.
Learning-based fluence map prediction offers a fast alternative to iterative inverse planning in intensity-modulated radiation therapy (IMRT), but its robustness under realistic distribution shifts remains unclear. We study a two-stage transformer pipeline that maps anatomy (CT and contours) to dose and then to beamlet fluence maps. We compare fluence-stage transformer backbones with hierarchical, global, and hybrid attention, trained with a physics-informed loss enforcing energy consistency. Robustness is evaluated under geometric perturbations, radiometric noise, reduced training data, and domain shifts using a prostate IMRT dataset, with additional evaluation of the dose stage on public datasets. Results show smooth degradation under moderate perturbations but sharp failures under severe rotations and noise. Hierarchical transformers (e.g., SwinUNETR) exhibit slower growth in upper-quartile energy error, indicating improved robustness. We further show that SSIM alone fails to capture clinically relevant errors, highlighting the need for physics-informed evaluation.