Next-level lung imaging: Darkfield X-ray, spectral CT, and AI-based reconstruction and analysis
The Fellowship advanced the clinical translation of dark-field radiography, spectral and photon-counting CT, and AI-based reconstruction to improve lung disease diagnosis e.g., for emphysema, pneumonia, and pneumothorax. We developed methods for dose reduction, quantitative biomarkers, artifact reduction, and lung-volume estimation using deep learning.
Focus Group: X-Ray and Computed Tomography Research
Prof. Daniela Pfeiffer (TUM University Hospital Rechts der Isar), Alumna Albrecht Struppler Clinician Scientist Fellow (funded as part of the Excellence Strategy of the federal and state governments)
Johannes Thalhammer (TUM), Doctoral Candidate
Collaboration Partners: Prof. Daniel Cremers, Prof. Franz Pfeiffer, Prof. Julia Herzen (TUM)
The Fellowship aimed to advance innovative X-ray-based imaging technologies for improved diagnosis, staging, and monitoring of lung disease. The scientific concept was built on three emerging imaging technologies: X-ray dark-field radiography and CT, spectral/photon-counting CT, and AI-driven reconstruction and analysis. This enabled us to achieve unprecedented sensitivity to lung microstructure, material composition, and functional surrogates. A unique element of the project was its access to the world’s first and, to date, only clinical dark-field chest radiography system for patients at TUM and provided a globally unique approach, allowing the project to investigate lung structure-function relationships and quantitative biomarkers from complementary physical, medical, and computational perspectives.
Summary of work carried out
The project advanced the clinical translation of dark-field chest radiography and CT, spectral CT methods, and deep-learning-based approaches to image reconstruction and biomarker extraction. The research activities included methodological developments in image acquisition, dose-efficient dark-field signal generation, quantitative scatter-based biomarkers, and AI-enhanced artifact reduction for sparse-view CT. While the original focus lay predominantly on pulmonary microstructure imaging, the scope expanded over time to include feasibility studies in musculoskeletal imaging. Outreach activities comprised invited conference talks, journal publications, interdisciplinary workshops, and participation in clinical feasibility studies with radiology partners.
A core achievement was demonstrating the diagnostic potential of X-ray dark-field imaging for structural lung disease. Urban et al. (Radiology 2022) and Urban et al. (Invest Radiol 2023) showed that dark-field chest radiography enables sensitive qualitative and quantitative assessment of pulmonary emphysema, providing one of the first clinical validations of dark-field signals as surrogates for microstructural tissue degradation and laying the groundwork for clinical translation (Fig. 1).
The Fellowship also advanced AI-based reconstruction and analysis. Thalhammer et al. (Radiol AI 2024) developed a U-Net-based artifact-reduction pipeline that improved hemorrhage detection in sparse-view CT, supporting dose-reduction strategies. Dorosti et al. (Radiol AI 2025) showed that deep-learning models can estimate total lung volume from single radiographs, enabling low-cost, widely deployable lung-function surrogates (Fig. 2).
A further translational highlight, Schaff et al. (Radiology 2024), demonstrated that dark-field radiography enhances detection of nondisplaced fractures often missed in conventional radiographs (Fig. 3), broadening the technology’s relevance beyond pulmonary imaging.
Figure 1
(C, D) in a healthy 33-year-old man (A, C) and a 65-year-old man with severe emphysema (B, D). The same window settings were applied within the respective modality. In C, no abnormalities are apparent. In D, flattened hemidiaphragms and an irregular area of radiolucency are visible. While the dark-field chest radiograph of the healthy subject exhibits a strong homogeneous dark-field signal (A), the dark-field signal intensity of the subject with pulmonary emphysema appears decreased overall and exhibits an inhomogeneous patchy pattern (B). Adapted from Urban et al. (Radiology 2022).
Major findings, outcomes, and impact
Our research delivered several high-impact findings. First, it established dark-field radiography as a clinically feasible, sensitive method for detecting microstructural lung impairment such as emphysema, pneumothorax, or early radiation-induced injury. These results confirm that dark-field provides information not accessible from conventional radiography, thus holding potential for earlier and more accurate diagnosis in chronic lung diseases.
Second, the Fellowship expanded the use of dark-field imaging beyond pulmonology. The feasibility of detecting nondisplaced fractures demonstrated its value in trauma diagnostics.
Third, the AI-based reconstruction work yielded practical improvements in CT imaging, particularly in scenarios requiring reduced dose or sparse-view acquisitions. The deep-learning-based artifact reduction significantly improved diagnostic performance, illustrating how physics-based imaging and modern computational methods can be synergistically integrated. Fourth, the project developed quantitative imaging biomarkers, including dark-field-based measures of microstructural integrity and deep-learning-derived lung-volume estimates. Together with spectral CT investigations into multimaterial decomposition and dual-energy biomarker extraction, the Fellowship contributed to a richer, clinically meaningful characterization of structure and function in diagnostic imaging.
Figure 2
Use of the final outcomes and future research directions
The outcomes have led to new clinical collaborations, strengthened international visibility, and laid the groundwork for ongoing projects in clinical dark-field translation, spectral CT evaluation, and AI-enabled quantitative lung imaging. The Fellowship thus significantly contributed to establishing the Fellow as a leading researcher at the intersection of advanced X-ray physics, medical imaging technology, and data-driven diagnostic innovation.
Future work will focus on expanding dark-field radiography into routine clinical pathways, particularly for COPD screening, therapy monitoring, and musculoskeletal applications. Combining dark-field with spectral CT and AI-based morphologic and functional analysis presents an exciting opportunity to develop hybrid biomarkers that jointly characterize microstructure, material composition, and ventilation surrogates. Additional research will address technical challenges such as workflow optimization, dose reduction, and large-scale validation. Further musculoskeletal and cardiovascular applications of dark-field imaging represent promising emerging directions. Finally, integrating generative AI and physics-informed neural networks into reconstruction and registration frameworks may substantially advance the diagnostic precision and efficiency of next-generation X-ray imaging.
Figure 3
Selected publications
- T. Urban et al., “Qualitative and Quantitative Assessment of Emphysema Using Dark-Field Chest Radiography,” Radiology, vol. 303, no. 1, pp. 119–127, Apr 2022, doi: 10.1148/radiol.212025.
- T. Urban et al., “Dark-Field Chest Radiography Outperforms Conventional Chest Radiography for the Diagnosis and Staging of Pulmonary Emphysema,” Invest Radiol, vol. 58, no. 11, pp. 775–781, Nov 1 2023, doi: 10.1097/RLI.0000000000000989.
- T. Dorosti et al., “Estimating Total Lung Volume from Pixel-Level Thickness Maps of Chest Radiographs Using Deep Learning,” Radiol Artif Intell, vol. 7, no. 4, p. e240484, Jul 2025, doi: 10.1148/ryai.240484.
- J. Thalhammer et al., “Improving Automated Hemorrhage Detection at Sparse-View CT via U-Net-based Artifact Reduction,” Radiol Artif Intell, vol. 6, no. 4, p. e230275, Jul 2024, doi: 10.1148/ryai.230275.
- F. Schaff et al., “Feasibility of Dark-Field Radiography to Enhance Detection of Nondisplaced Fractures,” Radiology, vol. 311, no. 2, p. e231921, May 2024, doi: 10.1148/radiol.231921.