DeepDRR – A Catalyst for Machine Learning in Fluoroscopy-Guided Procedures
Open accessThe DOI matches its Crossref record; 15 of the 19 references it cites were found in a registry or an index; 2 of the 12 sources its claims cite were checked for support.
Overview
CoChat Summary
TL;DR
DeepDRR is a framework for fast, realistic simulation of fluoroscopy and digital radiography from CT by combining learned material decomposition and scatter estimation with analytic forward projection and noise injection. Models trained on DeepDRRs generalized to unseen clinically acquired X-ray images for pelvic anatomical landmark detection without re-training or domain adaptation, whereas models trained on conventional DRRs failed.
Key findings
- Volumetric segmentation of air, soft tissue, and bone in CT achieved a misclassification rate of (2.03 ± 3.63)%.
- Scatter estimation on a 30-image test set yielded a mean squared error of 6.4% of the total scatter image energy.
- Simulating one 615 × 479-pixel X-ray projection took 2.0 s irrespective of the number of photons used.
- In pelvic landmark detection with 23 anatomical landmarks, a model trained on DeepDRRs produced accurate predictions on unseen clinical cadaver X-ray data, including partial anatomy, while the same model trained on conventional DRRs failed; no re-training or domain adaptation was performed.
Methods
- DeepDRR consists of four modules: deep ConvNet material decomposition in CT, material- and spectrum-aware ray-tracing forward projection, neural-network Rayleigh scatter estimation, and quantum/electronic readout noise injection.
- Material decomposition used an encoder-decoder ConvNet with skip connections, trained on 128 × 128 × 128-voxel patches with voxel sizes of 0.86 × 0.86 × 1.0 mm, multi-class Dice loss, 12 manually annotated whole-body CT volumes split into 10 training and 2 validation/testing, and 600 epochs.
- Scatter estimation used a ten-layer ConvNet trained on 330 Monte Carlo–generated images augmented by random rotations and reflections, with input downsampled to 128 × 128 pixels. Task-based evaluation used 23 pelvic landmarks, DRRs over a spherical segment covering 120° in RAO/LAO and 90° in CRAN/CAUD, 615 × 479-pixel images with 0.616² mm pixel size, a tungsten anode at 120 kV with 4.3 mm aluminum, and 5 × 10⁵ photons per pixel, tested on cadaver studies with a flat-panel C-arm system.
Limitations
- The paper states future work will focus on improving volumetric segmentation by introducing more materials, in particular metal, and on scatter estimation that could benefit from a larger training set size.
- Evident from the text: the task-based evaluation is a surrogate using cadaveric specimen data rather than live clinical patient data, and this paper reports qualitative landmark detection results while referring to [9] for comprehensive quantitative results.
- Evident from the text: material decomposition is limited to air, soft tissue, and bone.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
1 Valid · 2 Not valid · 9 Not checked · 9 Own results and interpretations
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“despite similar HU, the attenuation characteristic of bone is substantially different of soft tissue [10].”
What the sources say
- Could not verifycites [10] Hubbell, J.H., Seltzer, S.M.: Tables of X-ray mass attenuation coefficients and mass energy-absorption coefficients 1 keV to 20 MeV for elements Z = 1 to 92 and 48 additional substances of dosimetric interest. Technical report, National Institute of Standards and Technology (1995) sourceExperimental — accuracy not yet measured
Our copy of Hubbell & Seltzer (1995) is a data compilation: it tabulates and plots mass attenuation and mass energy-absorption coefficients (μ/ρ, μen/ρ) versus photon energy 1 keV–20 MeV, including separate entries for BONE, CORTICAL (ICRU-44) and TISSUE, SOFT (ICRU-44) (table/figure pages, e.g. the 'BLOOD, WHOLE (ICRU-44) / BONE, CORTICAL (ICRU-44) / BRAIN, GREY/WHITE MATTER (ICRU-44)' and 'TESTIS / TISSUE, SOFT (ICRU-44) / TISSUE, SOFT (ICRU FOUR-COMPONENT)' groupings), but nowhere does it state a comparison of bone versus soft tissue attenuation. Searches for 'Hounsfield unit / HU', 'bone cortical versus soft tissue attenuation differ', and 'effective atomic number / diagnostic radiology' returned no narrative passage — the only prose is definitional (mass thickness x = ρt, the cross-section decomposition σ = σpe+σcoh+σincoh+σpair+σtrip, Bragg-additivity departures for water vapor vs liquid). The claim's premise of 'similar HU' in particular has no counterpart in this source, and no passage contradicts it either, so the attribution cannot be checked against our copy. (LLM, conf=0.70)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“the model trained on conventional DRRs (upper row) fails to predict anatomical landmark locations on clinical data”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For scatter estimation, the evaluation on a test set consisting of 30 image yields a mean squared error of 6 . 4 % of the total scatter image energy.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For volumetric segmentation of air, soft tissue, and bone in CT volumes, we found a misclassification rate of (2 . 03 ± 3 . 63) %”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“the model trained on DeepDRRs produces accurate predictions even on partial anatomy”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
8 more claims
Create a free account to see all 21 claims and to verify the 10 references still open.
- It may state two separate claims as one. 5
- Probably not a substantial claim. 3
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Therefore, applying a synthetic data-trained machine learning model to real data is likely to result in substantially deteriorated performance (Unberath et al., 2018, 2019).”
From The Impact of Machine Learning on 2D/3D Registration for Image-Guided Interventions: A Systematic Review and Perspective (2021) · cites this paper as Unberath et al., 2018
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“A realistic forward projector that accounts for physically accurate image formation, while being capable of fast data generation, was used to obtain realistic synthetic training data. This allows direct application of the network model to real clinical X-ray images.”
From Learning to detect anatomical landmarks of the pelvis in X-rays from arbitrary views (2019) · cites this paper as [26]
Experimental — accuracy not yet measured
Citations
Cites
V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation
Milletari, F., Navab, N., & Ahmadi, S.-A. (2016) · 2016 Fourth International Conference on 3D Vision (3DV) · doi:10.1109/3dv.2016.79 · Cited by 11,931 · cited 2×
“63) % which is in line with results re-ported in previous studies using this architecture [3].”
BackgroundResultSelf-citationHubbell, J. H., & Seltzer, S. M. (1995) · doi:10.6028/nist.ir.5632 · Cited by 2,033
Large scale deep learning for computer aided detection of mammographic lesions
Kooi, T., Litjens, G., van Ginneken, B., Gubern-Mérida, A., Sánchez, C. I., Mann, R., den Heeten, A., & Karssemeijer, N. (2017) · Medical Image Analysis · doi:10.1016/j.media.2016.07.007 · Cited by 965 · cited 1×
“…by diagnostic imaging tasks where 1) all image data are archived, 2) learning targets, in particular annotations of any kind, exist traditionally [1] or can be approximated [2], and 3) comparably simple augmentation strategies, such as rigid and non-rigid displacements [3], ease the limited data…”
BackgroundSchneider, W., Bortfeld, T., & Schlegel, W. (2000) · Physics in Medicine and Biology · doi:10.1088/0031-9155/45/2/314 · Cited by 831
Badal, A., & Badano, A. (2009) · Medical Physics · doi:10.1118/1.3231824 · Cited by 358 · cited 4×
“The network was trained on 330 images generated via MC simulation [11], augmented by random rotations and reflections.”
BackgroundMethodInfluentialRussakoff, D. B., Rohlfing, T., Mori, K., Rueckert, D., Ho, A., Adler, J. R., & Maurer, C. R. (2005) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2005.856749 · Cited by 142 · cited 2×
“Ray-tracing algorithms are computationally efficient since the attenuated photon fluence of a detector pixel is determined by computing total attenuation along a 3D line that then applies to all photons emitted in that direction [8].”
BackgroundSisniega, A., Zbijewski, W., Badal, A., Kyprianou, I. S., Stayman, J. W., Vaquero, J. J., & Siewerdsen, J. H. (2013) · Medical Physics · doi:10.1118/1.4801895 · Cited by 117 · cited 2×
“A well studied example is variance reduction for scatter correction in cone-beam CT, since scatter is of low frequency [13].”
BackgroundMethodX-ray-transform Invariant Anatomical Landmark Detection for Pelvic Trauma Surgery
Bier, B., Unberath, M., Zaech, J.-N., Fotouhi, J., Armand, M., Osgood, G. M., Navab, N., & Maier, A. (2018) · Lecture notes in computer science · doi:10.1007/978-3-030-00937-3_7 · Cited by 113 · cited 3×
“To this end, we consider anatomical landmark detection in X-ray images of the pelvis from arbitrary views [9].”
MethodInfluentialSelf-citationFully Automatic and Real-Time Catheter Segmentation in X-Ray Fluoroscopy
Ambrosini, P., Ruijters, D., Niessen, W. J., Moelker, A., & van Walsum, T. (2017) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-66185-8_65 · Cited by 102 · cited 1×
“Despite clear opportunities, in particular for prediction tasks, very little work has considered learning in this context [4,5,6,7].”
BackgroundError Corrective Boosting for Learning Fully Convolutional Networks with Limited Data
Roy, A. G., Conjeti, S., Sheet, D., Katouzian, A., Navab, N., & Wachinger, C. (2017) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-66179-7_27 · Cited by 85 · cited 1×
“dominated by diagnostic imaging tasks where 1) all image data are archived, 2) learning targets, in particular annotations of any kind, exist traditionally [1] or can be approximated [2], and 3) comparably simple augmentation strategies, such as rigid and non-rigid displacements [3], ease the limited data problem.”
BackgroundSelf-citation
Cited by
A gentle introduction to deep learning in medical image processing
Maier, A., Syben, C., Lasser, T., & Riess, C. (2019) · Zeitschrift für Medizinische Physik · doi:10.1016/j.zemedi.2018.12.003 · Cited by 515
“Unberath et al. drive this even further to emulate the complete X-ray formation process in their DeepDRR [97].”
BackgroundGao, C., Killeen, B. D., Hu, Y., Grupp, R. B., Taylor, R. H., Armand, M., & Unberath, M. (2023) · Nature Machine Intelligence · doi:10.1038/s42256-023-00629-1 · Cited by 181
“As our experiments were precisely controlled and the only difference between the two scenarios is the image appearance due to varied simulation paradigm in the training set, this result supports the hypothesis that realistic simulation of X-rays using DeepDRR performs best for model transfer to real data.”
MethodInfluentialSelf-citationUnberath, M., Gao, C., Hu, Y., Judish, M., Taylor, R. H., Armand, M., & Grupp, R. (2021) · Frontiers in Robotics and AI · doi:10.3389/frobt.2021.716007 · Cited by 121
“…to align feature representations of real and synthetic data: [63, 34, 55] use heavy pixel-level transformations that approximate domain randomization, and [66, 45, 41, 63, 65, 70] rely on realistic synthesis to reduce domain shift, e.˙g., using open-source physics-based DRR engines [96, 89]).”
MethodSelf-citationX-ray-transform Invariant Anatomical Landmark Detection for Pelvic Trauma Surgery
Bier, B., Unberath, M., Zaech, J.-N., Fotouhi, J., Armand, M., Osgood, G. M., Navab, N., & Maier, A. (2018) · Lecture notes in computer science · doi:10.1007/978-3-030-00937-3_7 · Cited by 113
“We evaluate our approach on synthetic validation data and demonstrate that it generalizes to unseen clinical X-rays of the pelvis without the need of re-training.”
InfluentialSelf-citationEnabling machine learning in X-ray-based procedures via realistic simulation of image formation
Unberath, M., Zaech, J.-N., Gao, C., Bier, B., Goldmann, F., Lee, S. C., Fotouhi, J., Taylor, R., Armand, M., & Navab, N. (2019) · International Journal of Computer Assisted Radiology and Surgery · doi:10.1007/s11548-019-02011-2 · Cited by 63
“Nevertheless, the ten-layer ConvNet architecture as presented in [34] is heavily inspired by kernel-based methods but introduces modeling of a nonlinear relationship between the anatomy and scatter signal.”
InfluentialSelf-citation
Metadata
- Authors
- Unberath, M., Zaech, J.-N., Lee, S. C., Bier, B., Fotouhi, J., Armand, M., & Navab, N.
- Affiliations
- Johns Hopkins University (US)
- Corresponding author
- Unberath, Mathias
- Venue
- Lecture Notes in Computer Science
- Publisher
- Springer International Publishing
- Published
- 2018
- Pages
- 98-106
- ISSN
- 0302-9743 (ISSN-L)
- 1611-3349
- OpenAlex
- W2789356647
- Cited by
- 145
- References
- 19
- Citations by year
- 2026: 14 · 2025: 27 · 2024: 24 · 2023: 25 · 2022: 14 · 2021: 17 · 2020: 16 · 2019: 6 · 2018: 2
- FWCI
- 22.47
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- www.springernature.com/gp/researchers/text-and-data-mining (tdm)
- www.springernature.com/gp/researchers/text-and-data-mining (vor)
- Locations
- Lecture notes in computer science (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Engineering › Biomedical Engineering › Medical Imaging and Analysis
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Radiation Dose and Imaging
- Keywords
- material decomposition, scatter estimation, noise injection, synthetic training data, anatomical landmark detection, pelvis, interventional radiology, domain generalization, 2D-3D registration
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: DeepDRR – A Catalyst for Machine Learning in Fluoroscopy-Guided Procedures. https://app.cochat.ai/papers/10.1007/978-3-030-00937-3_12
- The paper
- Unberath, M., Zaech, J.-N., Lee, S. C., Bier, B., Fotouhi, J., Armand, M., & Navab, N. (2018). DeepDRR – A Catalyst for Machine Learning in Fluoroscopy-Guided Procedures. Lecture Notes in Computer Science. https://doi.org/10.1007/978-3-030-00937-3_12
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.