FluenceFormer: Transformer-Driven Multi-Beam Fluence Map Regression for Radiotherapy Planning

TL;DR

FluenceFormer addresses the ill-posed mapping from patient anatomy to deliverable fluence maps in IMRT planning by introducing a backbone-agnostic transformer framework with a two-stage design: Stage 1 regresses a global dose prior from anatomy, and Stage 2 conditions this prior on beam geometry to produce geometry-aware fluence maps. The Fluence-Aware Regression (FAR) loss combines voxel-level fidelity, gradient smoothness, structural correlation, and energy conservation to enforce physical realism, yielding deliverable plans. Evaluations across four transformer backbones on 99 prostate IMRT cases show that FluenceFormer, especially with Swin UNETR, achieves state-of-the-art energy accuracy (~4.5%) and improved structural fidelity (SSIM ≈ 0.70, PSNR ≈ 18.1) with statistically significant gains over CNN baselines and single-stage methods. The results demonstrate architecture-agnostic efficacy, highlight the value of intermediate dosimetric priors, and suggest practical impact for faster, more reproducible automated radiotherapy planning, while noting limitations related to cross-institution generalization and the potential for integrating differentiable dose calculation layers.

Abstract

Fluence map prediction is central to automated radiotherapy planning but remains an ill-posed inverse problem due to the complex relationship between volumetric anatomy and beam-intensity modulation. Convolutional methods in prior work often struggle to capture long-range dependencies, which can lead to structurally inconsistent or physically unrealizable plans. We introduce \textbf{FluenceFormer}, a backbone-agnostic transformer framework for direct, geometry-aware fluence regression. The model uses a unified two-stage design: Stage~1 predicts a global dose prior from anatomical inputs, and Stage~2 conditions this prior on explicit beam geometry to regress physically calibrated fluence maps. Central to the approach is the \textbf{Fluence-Aware Regression (FAR)} loss, a physics-informed objective that integrates voxel-level fidelity, gradient smoothness, structural consistency, and beam-wise energy conservation. We evaluate the generality of the framework across multiple transformer backbones, including Swin UNETR, UNETR, nnFormer, and MedFormer, using a prostate IMRT dataset. FluenceFormer with Swin UNETR achieves the strongest performance among the evaluated models and improves over existing benchmark CNN and single-stage methods, reducing Energy Error to $\mathbf{4.5\%}$ and yielding statistically significant gains in structural fidelity ($p < 0.05$).

Figures (3)

• Figure 1: Overview of the FluenceFormer framework. Stage 1 performs slice-wise dose regression from anatomical inputs. Stage 2 conditions the predicted dose on explicit geometric encoding (sin/cos) to resolve beam directionality and regress beam-specific fluence maps. ReLU-based regression heads replace segmentation classifiers to enable the prediction of unbounded physical intensities (Monitor Units).
• Figure 2: Qualitative comparison of fluence predictions for Beam 1, Beam 4, and Beam 5 across FluenceFormer and baseline methods.
• Figure 3: Component-level ablation of the FAR loss. Each variant alters the correlation and energy computation scopes. The beam-wise configuration (B/B) yields the highest PSNR and SSIM, indicating more stable and physically consistent fluence predictions.