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

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

Problem Statement
Learning-based fluence map prediction offers a promising alternative to iterative inverse planning in Intensity-Modulated Radiation Therapy (IMRT), potentially reducing planning time and inter-operator variability. However, the robustness of these models under realistic distribution shifts encountered during clinical deployment remains unclear. While conventional planning manages uncertainties (imaging noise, patient setup errors, scanner variability) through geometric margins and robustness constraints, the interaction between these uncertainties and learned representations in deep learning models is less transparent. Existing evaluations often prioritize accuracy on stable test sets, implicitly assuming data distributions that rarely hold in practice. Furthermore, fluence maps present a distinct risk profile: as dense regression targets constrained by beam geometry and energy conservation, small perturbations can induce non-linear modulation changes that are not captured by standard image-similarity metrics but result in clinically significant dosimetric deviations.

Methodology
The authors evaluate the robustness of a two-stage transformer-based fluence prediction pipeline, consistent with the FluenceFormer framework:

1. Anatomy-to-Dose Stage ($g_\phi$): Maps patient anatomy (CT and contours) to a predicted dose distribution.
2. Dose-to-Fluence Stage ($f_\theta$): Regresses beamlet fluence maps from the predicted dose and beam geometry.

The study focuses on the fluence stage ($f_\theta$) as it directly governs deliverable modulation, while also stress-testing the anatomy-to-dose stage on external datasets. The authors compare several transformer encoder backbones—SwinUNETR (hierarchical windowed attention), UNETR (fully global attention), nnFormer, and MedFormer (hybrid attention)—all trained with a Fluence-Aware Regression (FAR) loss. This loss function combines voxel-wise MSE, gradient consistency, correlation, and a physics-informed energy term to enforce global energy consistency.

Robustness is evaluated across four clinically motivated perturbation scenarios using a single-institution prostate IMRT dataset (99 patients) and two public datasets (OpenKBP head-and-neck and CORT prostate) for cross-dataset stress testing:

• Geometric Perturbations: Translational shifts (3–5 mm) and in-plane rotations (2°–5°).
• Radiometric Perturbations: Additive Gaussian noise ($\sigma = 0.05–0.20$) and smooth multiplicative bias fields.
• Data Scarcity: Training on reduced fractions of the dataset (25%, 50%, 75%).
• Domain Shifts: Synthetic domain shifts mimicking scanner variability.

Performance is quantified using the Structural Similarity Index Measure (SSIM) and, crucially, the 75th percentile of energy error ($EQ^{75}_{err}$) to capture tail-dose stability and physically meaningful failure modes.

Key Results

• Geometric Robustness: Rotational errors proved more disruptive than translational shifts. While all models showed smooth performance decay, SwinUNETR (hierarchical attention) exhibited the slowest growth in upper-quartile energy error ($EQ^{75}_{err}$) compared to fully global (UNETR) and hybrid variants. CNN baselines showed similar SSIM but significantly higher tail energy errors.
• Radiometric Robustness: All backbones demonstrated monotonic SSIM decay with increasing noise. nnFormer was the most sensitive (dropping >20% SSIM), whereas SwinUNETR and MedFormer were more tolerant, likely due to multi-scale representations suppressing high-frequency noise.
• Data Efficiency: All models improved with more data, but scaling behaviors differed. SwinUNETR remained the most data-efficient, achieving the highest SSIM and lowest $EQ^{75}_{err}$ even with only 25% of the training data. Conversely, nnFormer was highly data-dependent.
• Domain Robustness: Under bias-field perturbations, models like CNN and Swin-Direct maintained high SSIM but incurred large tail energy errors, indicating visually plausible but dosimetrically unsafe predictions. FluenceFormer variants, particularly SwinUNETR trained with physics-informed loss, better controlled energy drift.
• Cross-Dataset Generalization: The anatomy-to-dose stage generalized reasonably well to the OpenKBP and CORT datasets without fine-tuning, with DVH curves degrading smoothly under moderate perturbations.

Significance and Claims
The paper claims that robustness in fluence prediction is not solely a function of architectural novelty but emerges from the interplay between architectural inductive bias (specifically hierarchical windowed attention) and physics-informed supervision.

Key takeaways include:

1. Limitations of SSIM: Structural similarity metrics alone are insufficient for clinical safety assessment, as they can mask physically inconsistent predictions (e.g., high SSIM with large energy errors).
2. Importance of Energy Metrics: Energy-based metrics (like $EQ^{75}_{err}$) are critical for identifying failure modes that are clinically relevant but invisible to image-similarity measures.
3. Architectural Resilience: Hierarchical attention mechanisms (e.g., SwinUNETR) provide superior resilience to misalignment and noise compared to fully global attention, likely by limiting the propagation of spatial errors.
4. Deployment Safety: Systematic stress testing under realistic uncertainty, rather than clean-condition accuracy, is a prerequisite for the responsible clinical adoption of learning-based fluence prediction. The study delineates operating ranges where direct fluence prediction remains reliable and identifies regimes where robustness breaks down.