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

Ujunwa Mgboh, Rafi Sultan, Dongxiao Zhu, Joshua Kim · 2025

Read Fluence Map Prediction with Deep Learning: A Transformer-based Approach by Ujunwa Mgboh, Rafi Sultan, Dongxiao Zhu, Joshua Kim (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-08645.txt.

Introduction

Ujunwa Mgboh and three co-authors automate a slow step in prostate cancer radiotherapy: drawing the fluence maps that shape radiation beams. In intensity-modulated therapy these maps decide how much dose hits the tumor versus healthy tissue, and hand optimization depends heavily on planner skill.

A 3D Swin-UNETR network, a transformer design, reads CT scans and anatomy outlines and predicts nine-beam fluence maps directly. Training used ninety-nine prostate cases, with twenty held out for testing. Predicted maps, recalculated in a clinical planning system, correlate very strongly with the originals and pass gamma checks at high rates. Dose-volume measures show no significant differences from clinical plans. The authors present the system as a fast, consistent route to automated planning.

Catalog record

CREATORUjunwa Mgboh, Rafi Sultan, Dongxiao Zhu, Joshua Kim
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-08645

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Fluence Map Prediction with Deep Learning: A Transformer-based Approach Ujunwa Mgboh, Rafi Sultan, Dongxiao Zhu, Joshua Kim Abstract Background: Accurate fluence map prediction is a critical step in intensity-modulated radiation therapy (IMRT), where the linear accelerator output is modulated to maximize tumor coverage while minimizing dose to adjacent healthy tissues. Fluence map optimization remains one of the most time-consuming and expertise-dependent components of radiotherapy treatment planning. To address this, we propose a novel deep learning framework that accelerates fluence map generation while ensuring consistent and high-quality treatment plans. Methods: We developed an end-to-end 3D Swin-UNETR that directly predicts nine-beam fluence maps from volumetric CT images and anatomical contours. This transformer-based network leverages hierarchical self-attention to capture both local anatomical structures and long-range spatial dependencies. The dataset comprised 99 prostate IMRT cases (79 for training, 20 for testing). Predicted fluence maps were imported into the Eclipse Treatment Planning System (Varian Medical Systems, Palo Alto, CA) for multileaf collimator sequencing and dose recalculation. Model performance was assessed using beam-wise fluence correlation, spatial gamma analysis, and DVH metrics. Results: The produced fluence maps yielding dose distributions in strong agreement with clinical reference plans. On the 20-patient test set, it achieved an average R 2 = 0.95 ± 0.02 R^{2}=0.95\pm 0.02 , MAE = 0.035 ± 0.008 \mathrm{MAE}=0.035\pm 0.008 , and gamma passing rates of 85 ± 10 % 85\pm 10\% (3%/3 mm). No statistically significant differences were observed in DVH parameters for target volumes or organs at risk. The proposed Swin-UNETR framework enables fully automated fluence map prediction directly from anatomical inputs, eliminating reliance on precomputed dose information. By exploiting hierarchical self-attention for volumetric context modeling, the method enhances accuracy, spatial coherence, and efficiency in IMRT plan generation. These findings highlight its potential as a clinically scalable, inverse-free solution for automated radiotherapy planning. 1 Introduction Radiotherapy uses high-energy radiation to target and kill cancer cells. The goal in all of radiation therapy treatment planning is to optimize the delivered dose to administer a tumorcidal dose of radiation to the targeted lesion while limiting the amount of dose given to surrounding healthy tissues. Within the field of radiation oncology, the most time-consuming step in the radiotherapy planning workflow has historically been the optimization of the treatment plan, which is an iterative process requiring several trials before the planner can achieve an optimal plan that best matches the radiation oncologist’s plan goals. Recently, the implementation of deep learning models in the treatment planning workflow has emerged as a potentially effective method for dramatically reducing the time required for plan generation and standardizing fluence map prediction accuracy and effectiveness, leveraging techniques used in fields like natural language processing and computer vision r1 ; r2 ; r3 ; r4 Various input features are utilized in deep learning models for treatment plan automation tasks, influenced by factors like the specific radiotherapy application, input availability, and model selection. For instance, r7 enhanced gamma passing rate prediction and classification by combining dosiomics and plan features, leading to improved outcomes. Similarly, in predicting SBRT outcomes, a combination of dose distribution images and patient data enabled the identification of post-SBRT occurrence likelihood. In the context of fluence map prediction, input types such as image-based, dosimetric features, or machine learning-generated data are employed, aiming to optimize efficiency and accuracy in automated radiotherapy treatment planning r9 ; r10 . Fluence map prediction, traditionally a labor-intensive process, has seen advancements through automated treatment planning (ATP) using deep learning models. Notably, a double convolutional neural network (CNN) incorporating patient anatomy and beam templates was employed to predict fluence maps for pancreas stereotactic body radiation therapy r5 . Similarly, sequential CNN architectures tailored to unique input features and model designs for fluence prediction has been introduced r12 ; r13 ; r14 . Related studies primarily employed ResNet or UNet-based architectures. ResNet variants were used to enhance prediction accuracy r12 and to integrate knowledge-based planning for improved plan features r6 . UNet-based models, including both 2D and 3D variants, were applied for dose and fluence map prediction r14 , as well as for mapping dose distributions to fluence maps with demonstrated generalizability across cancer sites and radiotherapy techniques r17 ; r16 . These experiments further revealed insights into model performance concerning plan quality and delivery modalities, highlighting the potential of deep learning in fluence map prediction for radiotherapy planning. The ability of deep learning to directly predict fluence/intensity maps without going through the time-consuming and disparity-prone inverse mapping process has been explored using multiple deep learning models r5 ; r6 . Building upon previous frameworks, this study extends fluence map prediction toward a comprehensive and clinically deliverable IMRT planning workflow for prostate cancer. We created a deep learning model from start to finish using the Swin UNETR architecture. This model uses hierarchical self-attention to pick up on both fine-grained anatomical details and long-range spatial dependencies in volumetric data. The model takes CT images and anatomical contours as inputs and uses them to predict nine volumetric fluence maps that match the beam intensities that can be delivered. The proposed framework learns the implicit relationship between anatomy and fluence within the network itself, eliminating the need for any external dose input. The predicted fluences are then imported into the treatment planning system for final dose calculation and DVH evaluation, the proposed method demonstrates the effectiveness of transformer-based architectures for accurate and clinically interpretable fluence prediction in radiotherapy planning. 2 Materials and Methods 2.1 Dataset Details In this research work, we utilized datasets from ninety-nine (99) prostate patients who underwent radiation therapy at our institution. Computed Tomography (CT) images were acquired using a Brilliance Big Bore (Philips Health Care, Cleveland, OH) scanner with the following parameters: 140 kVp, 500 mAs, 512x512 in-plane image dimensions, 1.28x1.28 {mm}^{2} in-plane spatial resolution, and 3 mm slice thickness. Patients were originally treated using IMRT plans with 9 equally spaced fields to a total dose of 70.2-79.2 Gy in 1.8-2.0 Gy/fx. All plans were generated in the Eclipse TPS. Standard dose-volume histogram (DVH) metrics were utilized for target and organs-at-risk (OAR). Bladder, prostate, left and right femoral heads, and penile bulb were evaluated using criteria defined in QUANTEC guidelines r26 . Differences between the dose resulting from the predicted fluences and the original dose were evaluated using a paired two-sample t-test with a significance level of p<0.05 . 2.2 Dataset Preprocessing The CT images, anatomical contours, and dose distributions for all patients were exported from the Eclipse Treatment Planning System (TPS) as DICOM files. The CT and contour data were processed using MATLAB (MathWorks, Natick, MA) scripts to generate structured arrays for each patient, while the fluence maps originally stored as text-based “.optimal fluence” files were extracted and converted into NumPy arrays for compatibility with the deep learning framework. All images were then resized to 128\times 128 pixels to maintain uniform spatial resolution across patients and modalities. Input and Output Representation Each patient case was represented as a multi-channel volumetric tensor that integrates both anatomical and structural information. The CT and contour volumes were concatenated to form the model’s input tensor, allowing the network to jointly capture spatial relationships between anatomy and treatment structures. The model was trained in an end-to-end fashion to predict nine volumetric

Read the complete text