Uncertainty-guided Generation of Dark-field Radiographs
Lina Felsner, Henriette Bast, Tina Dorosti, Florian Schaff, Franz Pfeiffer, Daniela Pfeiffer, Julia Schnabel · 2026
Read Uncertainty-guided Generation of Dark-field Radiographs by Lina Felsner, Henriette Bast, Tina Dorosti, Florian Schaff, Franz Pfeiffer, Daniela Pfeiffer, Julia Schnabel (2026) 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-2601-15859.txt.
Introduction
Lina Felsner and six co-authors make a new kind of X-ray image in a 2026 arXiv paper. X-ray dark-field imaging shows fine tissue structure in a way plain X-rays do not. It gives clues that standard scans miss. But dark-field data are scarce. That scarcity hampers deep learning models, which need large datasets. The team builds the first frame of its kind. It makes dark-field images from standard chest X-rays.
Their tool is an uncertainty-guided progressive GAN. Maps of doubt steer the build toward detail that can be trusted. Growth in stages builds the image step by step. Fake dark-field data could unblock model work in this young field. Hospitals cannot easily collect paired scans. A sound tool gives researchers data to train and test methods now.
Catalog record
| CREATOR | Lina Felsner, Henriette Bast, Tina Dorosti, Florian Schaff, Franz Pfeiffer, Daniela Pfeiffer, Julia Schnabel |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 21 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-15859
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Uncertainty-guided Generation of Dark-field Radiographs Lina Felsner, Henriette Bast, Tina Dorosti, Florian Schaff, Franz Pfeiffer, Daniela Pfeiffer, Julia Schnabel Abstract X-ray dark-field radiography provides complementary diagnostic information to conventional attenuation imaging by visualizing microstructural tissue changes through small-angle scattering. However, the limited availability of such data poses challenges for developing robust deep learning models. In this work, we present the first framework for generating dark-field images directly from standard attenuation chest X-rays using an Uncertainty-Guided Progressive Generative Adversarial Network. The model incorporates both aleatoric and epistemic uncertainty to improve interpretability and reliability. Experiments demonstrate high structural fidelity of the generated images, with consistent improvement of quantitative metrics across stages. Furthermore, out-of-distribution evaluation confirms that the proposed model generalizes well. Our results indicate that uncertainty-guided generative modeling enables realistic dark-field image synthesis and provides a reliable foundation for future clinical applications. 1 School of Computation, Information and Technology, Technical University of Munich, 85748 Garching, Germany 2 Munich Center for Machine Learning (MCML), Munich, Germany 3 Institute of Machine Learning in Biomedical Imaging, Helmholtz Munich, 85764 Neuherberg, Germany 4 Department of Physics, School of Natural Sciences, Technical University of Munich, 85748 Garching, Germany 5 Munich Institute of Biomedical Engineering, Technical University of Munich, 85748 Garching, Germany 6 Institute for Diagnostic and Interventional Radiology, School of Medicine and Health, TUM University Hospital Klinikum rechts der Isar, Technical University of Munich, 81675 Munich, Germany 7 Institute for Advanced Study, Technical University of Munich, 85748 Garching, Germany 8 School of Biomedical Engineering & Imaging Sciences, King’s College, London, UK Keywords: X-ray dark-field image generation, Generative adversarial networks, Aleatoric and epistemic uncertainty modeling 1 Introduction X-ray dark-field radiography is an advanced imaging modality that complements conventional X-ray by revealing microstructural features via small-angle scattering [ 11 , 17 ] . The Dark-field signal can be measured jointly with an attenuation image via grating-based interferometer setups [ 11 , 16 ] (see Fig. 1 ). Unlike standard attenuation-based radiography, where dense structures such as ribs and the heart dominate, dark-field contrast originates from microscopic interfaces that scatter X-rays. The numerous alveolar air–tissue interfaces in healthy lungs produce a strong dark-field signal making healthy lungs appear bright, while diseases that affect alveoli (e.g. emphysema) cause patchy dark regions due to loss of scattering [ 16 , 15 ] . Early clinical studies indeed suggest that dark-field chest radiography provides unique diagnostic value to quantify pulmonary emphysema in COPD patients [ 16 , 15 ] and enhance COVID-19 diagnosis [ 7 ] . Deep learning has shown transformative potential in medical imaging, achieving strong performance in tasks such as nodule detection, disease classification, and image segmentation on conventional radiographs [ 19 ] . However, robust training of such models typically demands large, diverse datasets, a requirement that is challenging for a new modality like dark-field, where only limited human data currently exist. One promising strategy to mitigate this problem is synthetic data generation: using generative models to create realistic artificial images for data augmentation [ 3 , 18 , 2 , 14 , 10 ] . By augmenting the training set with high-fidelity simulated samples [ 19 ] , or using synthetic pretraining [ 9 ] model performance can be improved. \begin{overpic}[figures/setup3.png] \put(0.0,23.5){\color[rgb]{0,0,0}a} \put(43.0,23.5){\color[rgb]{1,1,1}b} \put(74.0,23.5){\color[rgb]{1,1,1}c} \end{overpic} Figure 1: (a) Measurement setup for the grating-based dark-field scanner. The subject stands in the scanner while the grating assembly scans from bottom to top. (b) Attenuation image and (c) dark-field image of a healthy subject, jointly acquired in a single scan. Generative Adversarial Networks (GANs) have been especially influential because they can synthesize realistic images via adversarial training without explicit density modeling, and tend to require fewer training samples than diffusion models. Many GAN architectures have been explored for medical imaging tasks, from MedGAN [ 2 ] , a general-purpose medical image translation framework operating end-to-end on image inputs, to task-specific models like PathologyGAN [ 12 ] , which learns to generate and analyze realistic histopathology images of cancer tissue. Overall, the medical imaging field is increasingly leveraging such synthetic data techniques to supplement limited datasets and improve AI model training [ 5 ] . An essential aspect for the clinical deployment of generative models is the quantification of uncertainty. Uncertainty generally falls into two categories: aleatoric uncertainty, which reflects noise inherent to the observed data, and epistemic uncertainty, representing the uncertainty arising from the model itself—this latter type can be reduced with sufficient data [ 8 ] . Uncertainty quantification techniques are critical for mitigating uncertainties in both optimization and decision-making processes [ 1 ] . Recent research indicates that incorporating uncertainty estimation not only enhances interpretability but can also lead to improved model performance [ 14 ] . For instance, Upadhyay et al. [ 14 ] introduced an uncertainty-guided GAN approach for image translation trained in a cascaded matter. They show that the visual quality and quantitative metrics improve across these refinement phases, and that the uncertainty maps help the model focus on areas of poor synthesis. Moreover, the uncertainty estimation improves interpretability and may support expert assessment and safe deployment in medical contexts [ 14 ] . These approaches highlight that developing reliable medical image synthesis systems requires not only generating realistic images but also identifying low-confidence regions to mitigate false certainty. In this paper, we propose to use an uncertainty-guided image synthesis framework to generate X-ray dark-field images from standard attenuation chest X-rays. To our knowledge, this is the first attempt to translate conventional radiographs into dark-field contrast via a generative model, enabling the large-scale generation of virtual dark-field data from widely available chest X-rays. The main contributions of this work can be summarized as follows: 1. To the best of our knowledge we are the first to generate dark-field images from conventional 2D X-ray radiographs. 2. We explicitly model both aleatoric and epistemic uncertainty within the image generation process to improve interpretability and reliability. 3. We evaluate the proposed model on the ChestX-ray dataset to assess its robustness and generalization under domain-shift conditions. The proposed framework not only provides a new tool for generating synthetic dark-field datasets to facilitate machine learning, but also moves toward reliable AI-assisted interpretation by quantifying uncertainty in the generated images. Together, these contributions aim to accelerate the use of dark-field imaging for lung disease diagnosis by combining image generation with principled uncertainty estimation. 2 Methods Figure 2: Proposed Uncertainty-Guided Progressive GAN framework, illustrated for model training at stage three. At this stage, the previous layers are frozen. The aleatoric uncertainty estimates are used as attention maps to guide the model’s weights for refinement in the subsequent stage. Dropout is incorporated to the generators at each stage. We propose a generative approach to synthesize dark-field contrast images from conventional attenuation chest radiographs using a GAN-based framework that incorporates both aleatoric and epistemic uncertainty. Our method builds on the Uncertainty-Guided Progressive GAN framework introduced by Upadhyay et al. [ 14 ] , based on a progressive learning scheme and uses aleatoric uncertainty estimates as attention maps to guide the models weight for refinement. In this work, we explicitly model both forms of uncertainty: (i) Aleatoric uncertainty, which accounts for the inherent noise in the observations, and (ii) Epistemic uncertainty, which arises from limited knowledge