Robustness of Transformer-Based Fluence Map Prediction Under Clinically Realistic Perturbations

TL;DR

This study evaluates the robustness of transformer-based models for fluence map prediction in radiation therapy under realistic perturbations, highlighting the importance of physics-informed metrics.

Contribution

It introduces a physics-informed loss for training transformers and compares different attention mechanisms, demonstrating improved robustness in clinical scenarios.

Findings

Hierarchical transformers show slower error growth under perturbations.

Severe rotations and noise cause sharp failures in predictions.

SSIM alone is insufficient to detect clinically relevant errors.

Abstract

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.