Fluence Map Prediction with Deep Learning: A Transformer-based Approach

TL;DR

This work tackles the slow, planner-dependent process of IMRT fluence map generation by introducing an end-to-end 3D Swin-UNETR transformer that maps volumetric CT images and anatomical contours directly to nine deliverable fluence maps for prostate cancer. Trained on 99 prostate cases and evaluated on 20 unseen cases, the model achieves strong dosimetric and spatial agreement with clinical plans, yielding $R^2 = 0.95 \pm 0.02$, $MAE = 0.035 \pm 0.008$, $RMSE = 0.050 \pm 0.010$, and a gamma pass rate of $85\% \pm 10\%$ (3%/3 mm), while processing each patient in roughly $3.97$ seconds. The framework preserves clinically relevant DVH metrics with no significant differences from reference plans, demonstrating potential to substantially accelerate automated IMRT planning. Limitations include prostate-only data and a single-institution evaluation, motivating future multi-institutional and multi-site validation to establish generalizability and scalability.

Abstract

Accurate fluence map prediction is essential in intensity-modulated radiation therapy (IMRT) to maximize tumor coverage while minimizing dose to healthy tissues. Conventional optimization is time-consuming and dependent on planner expertise. This study presents a deep learning framework that accelerates fluence map generation while maintaining clinical quality. An end-to-end 3D Swin-UNETR network was trained to predict nine-beam fluence maps directly from volumetric CT images and anatomical contours using 99 prostate IMRT cases (79 for training and 20 for testing). The transformer-based model employs hierarchical self-attention to capture both local anatomical structures and long-range spatial dependencies. Predicted fluence maps were imported into the Eclipse Treatment Planning System for dose recalculation, and model performance was evaluated using beam-wise fluence correlation, spatial gamma analysis, and dose-volume histogram (DVH) metrics. The proposed model achieved an average R^2 of 0.95 +/- 0.02, MAE of 0.035 +/- 0.008, and gamma passing rate of 85 +/- 10 percent (3 percent / 3 mm) on the test set, with no significant differences observed in DVH parameters between predicted and clinical plans. The Swin-UNETR framework enables fully automated, inverse-free fluence map prediction directly from anatomical inputs, enhancing spatial coherence, accuracy, and efficiency while offering a scalable and consistent solution for automated IMRT plan generation.

Figures (3)

• Figure 1: Model architecture. Overview of the proposed end-to-end deep learning framework for radiotherapy fluence map prediction. The model integrates volumetric CT images and contours into a unified input tensor and employs a 3D Swin-UNETR architecture to directly predict nine deliverable fluence maps corresponding to the clinical beam angles. The Swin-UNETR backbone combines hierarchical self-attention and UNet-style skip connections to capture both fine-grained anatomical structure and long-range spatial dependencies, enabling spatially coherent and clinically interpretable fluence predictions.
• Figure 2: Comparison of clinical and predicted dose distributions for representative test cases. Panels (a–c) show beam-level fluence maps for two beams (Beam 4 and Beam 5), including (a) the clinical plan, (b) the model-predicted plan, and (c) the voxel-wise difference (clinical $-$ predicted). Panels (d–f) display axial composite dose distributions: (d) clinical, (e) predicted, and (f) percentage dose difference ($\Delta$Dose, %). The predicted doses exhibit high spatial and dosimetric agreement with the clinical reference.
• Figure 3: Dose–volume histogram (DVH) for Patient 70. The overlap shows significant similarity between the ground truth and predicted plan.