Leveraging Convolutional Neural Networks for 3D Quantitative Angiography Reconstructions from Sparse Cone Beam CT Projections Utilizing CFD Data

Ahmad Rahmatpour; Allison Shields; Ciprian N Ionita; Kyle A Williams; Michael Udin; Mohammad Mahdi Shiraz Bhurwani; Parisa Naghdi; Parmita Mondal; Swetadri Vasan Setlur Nagesh

REVIEW 4 major objections 6 minor 7 references

Leveraging Convolutional Neural Networks for 3D Quantitative Angiography Reconstructions from Sparse Cone Beam CT Projections Utilizing CFD Data

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A modified U-Net can reconstruct 3D angiographic volumes from truncated cone-beam projections, achieving a test MSE of 1e-4 on CFD-simulated aneurysm cases.

desk verdict Useful CFD-based simulation pipeline, but evaluating against FDK instead of the true contrast field leaves the feasibility claim untested. read the letter →

arxiv 2411.09632 v2 pith:KYYIKPDC submitted 2024-11-14 physics.med-ph

Ahmad Rahmatpour , Allison Shields , Parmita Mondal , Parisa Naghdi , Michael Udin , Kyle A Williams , Mohammad Mahdi Shiraz Bhurwani , Swetadri Vasan Setlur Nagesh

show 1 more author

Ciprian N Ionita

This is my paper · ORCID

classification physics.med-ph

keywords 3DangiographyintracranialaneurysmssparseprojectionsconebeamCTcomputationalfluiddynamicsconvolutionalneuralnetworksU-Netquantitative

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that a modified U-Net can turn truncated cone-beam CT projection sets into full 3D angiographic volumes, using synthetic data as proof. The authors simulate contrast-enhanced blood flow in three patient-derived intracranial aneurysm geometries with computational fluid dynamics, project those volumes through a simulated C-arm CT system, and train the network to map reconstructions from at most 49 projections to Parker-weighted FDK reconstructions from 108 projections. On the test set the model reports a mean squared error of $10^{-4}$, which the authors read as evidence that transient contrast features survive the reconstruction. The clinical motivation is that halving the number of projections roughly doubles the temporal resolution of 3D quantitative angiography, letting clinicians isolate the brief arterial phase. The claim is deliberately scoped to simulation: the authors say further validation is needed before clinical implementation.

What carries the argument

The load-bearing object is a modified U-Net, an encoder-decoder convolutional network whose contracting path downsamples with stacked convolutional blocks (kernel sizes 8x8, 5x5, and 3x3; filter count doubling from 64) and ReLU activations, and whose expanding path upsamples and concatenates the corresponding encoder features to restore spatial detail. A final 1x1 convolution with linear activation produces an output volume matching the input size. The network is trained with the Adam optimizer and mean squared error loss to map truncated reconstructions to their fully sampled FDK counterparts, which serve as ground truth. The supporting pipeline that generates its training data is equally central: CFD simulations with labeled particles mimic contrast bolus transit, and a simulated cone-beam C-arm generates both the full 108-projection acquisitions and the truncated sets.

What would settle it

Retrain the model with one aneurysm geometry completely withheld from training and test on that geometry: if vessel-region error rises far above the reported $10^{-4}$, the network has memorized the three simulated shapes rather than learned a general truncated-to-full reconstruction map. A more direct test compares the network output to the true CFD contrast volume rather than to FDK; if the FDK-vs-CFD gap is as large as the reported error, the claimed fidelity is an artifact of matching one approximation to another.

Watch

Extended reading notes

Core claim

The authors claim that a modified U-Net convolutional network can reconstruct 3D quantitative angiographic volumes from sparse cone-beam projections with sufficient fidelity to preserve the time-varying contrast distribution in intracranial aneurysms. The supporting demonstration is entirely in silico: steady-flow CFD with particle-labeled contrast boluses, simulated C-arm acquisition over a 220-degree arc at 27 frames per second, Parker-weighted FDK volumes as ground truth, and truncated reconstructions from 49 (and, in augmented data, 29) projections as network input. A test-set MSE of $10^{-4}$ is presented as evidence that the network captures the relevant transient angiographic features, and the authors conclude that CNN-based reconstruction from truncated projections is feasible.

Load-bearing premise

The evaluation assumes that a Parker-weighted FDK reconstruction from 108 fully sampled simulated projections is an accurate ground truth for a 3D angiogram, and that the CFD particle-labeling simulation faithfully reproduces real contrast transport.

Editorial extensions

If this is right

If the reported accuracy holds in real acquisitions, a C-arm running the same 27 fps gantry motion could produce dynamic 3D volumes at roughly twice the frame rate, since only about half the projections are needed.
Clinicians could extract time-density (contrast dilution) curves at every voxel of the 3D volume, turning 2D parametric maps into 3D quantitative angiography of the aneurysm.
Radiation dose could be reduced proportionally to the projection truncation, because fewer angles are irradiated during the acquisition.
The result is currently a feasibility statement about matching FDK volumes, not about reconstructing anatomy from raw sparse projections, and the authors state that translation to clinical practice requires further validation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the ground truth is itself an FDK reconstruction from the same simulated projection geometry, the MSE measures how well the network reproduces an FDK approximation, not how close either volume is to the true contrast distribution; an extension would compare against the CFD volume directly.
With only three patient-derived geometries, the network could be memorizing shape-specific intensity patterns; a leave-one-geometry-out test would show whether the reconstruction generalizes to unseen vascular anatomy.
The steady-flow, particle-labeling contrast model omits pulsatile flow and rapid mixing, so the temporal features the network learns may not match real bolus dynamics; retraining on pulsatile CFD or physically measured phantoms would test this.
Since most of a 256x256x256 angiographic volume is background, the overall MSE of 10^-4 may be dominated by easy zero voxels; vessel-only or structure-aware metrics (e.g., Dice on segmentations) would reveal where the errors actually live.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a small synthetic proof-of-concept, and the central claim outruns the evidence. The one genuinely useful thing is the CFD-to-CBCT pipeline for generating dynamic 3D angiography data with known ground truth — that part is well thought out, with three patient-specific geometries, multiple inlet velocities and injection durations, and a realistic C-arm model. That pipeline could be reused by others.

The paper does not do the obvious validation. The true contrast concentration fields are known from CFD, but the authors never compare either their truncated FDK inputs or their full-angle FDK 'ground truth' to those fields. So the reported MSE of 1e-4 only says the U-Net can turn one reconstruction algorithm's output into another's. It says nothing about whether the network recovers actual contrast dynamics. That's a load-bearing gap, not a quibble. Also, the network maps volume to volume (truncated FDK to full FDK), so the abstract's phrasing about 'reconstruction from projections' overstates the setup.

Other soft spots are more typical for a conference-length paper: no baseline comparison (e.g., FDK alone or a simpler interpolation), no per-case or variance numbers, and no code or data release. Three geometries is small, but acceptable for a feasibility study. The citation pattern is fine; the self-citations are to the group's prior work on parametric imaging, and they are used appropriately.

Who is this for? Researchers working on sparse-view CBCT artifact reduction, especially for neurovascular imaging. They might get a useful data-generation recipe out of it. But the core claim needs revision. I'd like to see a comparison between both FDK reconstructions and the CFD truth, plus at least one classical baseline.

Would I send this to peer review? Yes, because the pipeline is worth referee time and the flaws are fixable with a moderate amount of additional experiment. A desk reject would lose the useful parts. But it needs major revisions before publication.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes a modified U-Net to reconstruct 3D angiographic volumes from sparse cone-beam CT data, with the goal of improving temporal resolution in intracranial aneurysm imaging. The authors generate three patient-derived vascular geometries, run CFD simulations of blood flow and contrast bolus transport, simulate a C-arm CBCT acquisition over a 220-degree arc, and reconstruct volumes with a Parker-weighted FDK algorithm. Truncated-angle FDK volumes are used as network input, and full-angle FDK volumes are used as training targets. The network is a standard U-Net with a 256x256x256 input and MSE loss. The authors report a test MSE of 1e-4 and conclude that the model can capture transient angiographic features, supporting the feasibility of CNN-based 3D reconstruction from truncated projections.

Significance. If the central claim were established, the work would be clinically relevant because it could enable 3D quantitative angiography with fewer projections, improving temporal resolution during contrast bolus passage. The in-silico pipeline that combines patient-specific CFD with simulated CBCT acquisition is a useful testbed and the paper deserves credit for clearly describing the data-generation machinery. However, as presented, the evidence supports only a narrower claim: a U-Net can map truncated-angle FDK volumes to full-angle FDK volumes on a small set of synthetic cases. The reported MSE is not compared with any baseline, no per-patient breakdown is given, and the ground truth itself is never validated against the known CFD contrast fields. Therefore the significance of the stated feasibility conclusion is currently not supported.

major comments (4)

[Section 2.2 and Section 3] The training targets are 'the corresponding fully reconstructed CT slices,' which are Parker-weighted FDK volumes from 108 simulated projections (Section 2.1). The reported MSE of 1e-4 therefore measures agreement between a truncated-angle FDK volume and a full-angle FDK volume, not agreement with the true CFD contrast concentration field. Because the simulation pipeline makes the true 3D angiographic volumes available, the authors should compare both FDK reconstructions and the network output against those true volumes. Without this comparison, the claim in Section 3 that the model 'was capable of capturing important transient angiographic features' is unsupported; the network may merely be learning to imitate FDK's artifact structure.
[Abstract and Section 2.2] The abstract and title state that the network reconstructs 3D angiographic volumes from 'sparse cone beam CT projections,' but the network input is a truncated FDK reconstruction (a 256x256x256 volume), not raw projection data. The described method is therefore a volume-to-volume artifact-reduction or regression task, not a projection-to-volume reconstruction. This distinction is not merely semantic: if the input were raw sparse projections, the network would need to handle the projection geometry, which the current architecture does not. The claims should be revised to match the actual experimental setup.
[Section 3] The quantitative evaluation is incomplete. The MSE of 1e-4 is reported without a baseline (e.g., the MSE of the input truncated FDK volume against the full FDK target, or the MSE of a constant-zero prediction). Since angiographic volumes are largely background with zero intensity, an uninformative model can achieve a very small MSE. In addition, no variance, confidence interval, or per-patient breakdown is provided, and the test set composition (which of the three geometries, how many volumes) is not described. Without these comparisons, the numerical result in Section 3 cannot be interpreted as evidence of accurate reconstruction.
[Section 3, Figure 4] The claim that the model captures 'transient angiographic features' is supported only by a qualitative figure showing reconstructed volumes at different times. No quantitative evaluation of temporal dynamics is provided, such as comparison of time-density curves between the network output, the FDK ground truth, and the CFD-simulated contrast concentrations. Without such analysis, the conclusion that the network accurately captures contrast progression over time is not demonstrated.

minor comments (6)

[Abstract/Summary] The abstract says 'truncated projection datasets... using a maximum of 49 projections,' while the Summary states the model was trained on 'sparse reconstructions from 29 and 49 projections.' Please clarify the actual number of projections used in the experiments.
[Figure 3] The caption says 'Phantoms generated by computational fluid dynamics from different angle projections,' but the figure appears to show reconstructions (truncated, ground truth). Please clarify what each panel displays and label the axes and angles consistently.
[Section 2.1] The text states that data were recorded every 10 ms, 'generating thousands of volumetric angiograms for each case,' but then describes CBCT acquisitions at 27 fps with 108 projections. Please clarify the relationship between the temporal sampling of the CFD volumes and the projection acquisition rate.
[Section 2.1] The phrase 'high-fidelity volumetric images' overstates the quality of Parker-weighted FDK reconstructions from a 220-degree arc; please use more neutral language such as 'full-angle FDK reconstructions.'
[References] Reference [6] is incomplete (no journal, volume, or pages) and reference [7] lacks the conference or proceedings title. Please complete the bibliographic entries.
[General] There are minor grammatical issues, such as 'data was recorded' (should be 'data were recorded') and inconsistent capitalization of 'an giographic' in the abstract. A careful proofreading pass is recommended.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported 10^-4 MSE is measured against FDK reconstructions that are also the training target, so the feasibility claim rests on a self-referential 'ground truth'.

self definitional [Section 2.2 (Model investigation) and Section 3 (Results)]
"The truncated reconstructions served as the training set, while the corresponding fully reconstructed CT slices were used as ground truth. ... Validation of the U-Net model on the test set yielded a MSE of 10-4. This indicated that the model performed well in accurately reconstructing the ground truth and was capable of capturing important transient angiographic features."

Ground truth is the Parker-weighted FDK output from the same simulated projection pipeline that produced the truncated inputs, and the U-Net was trained to minimize MSE to it. The reported 10^-4 MSE is computed against those same FDK volumes, while the actual CFD contrast fields (Section 2.1) are never used as benchmarks. The claim of capturing transient features is therefore supported only by agreement with the reconstruction algorithm that generated the training target; the metric is truncated-FDK-to-full-FDK regression, not recovery of true contrast dynamics.

full rationale

The forward CFD and simulated cone-beam CT pipeline is an independent simulation step and is not circular. The U-Net training on held-out simulated cases is also a genuine empirical procedure. The circularity is concentrated in the validation target: the paper calls Parker-weighted FDK reconstructions 'ground truth' and then presents agreement with those reconstructions as evidence of capturing real transient contrast features. Because the true CFD angiographic volumes were available and not used, the central feasibility claim reduces to the network matching the very reconstruction algorithm it was taught to imitate. The self-citations in the reference list are background and not load-bearing. Overall, the derivation is partially self-referential but not equivalent to its inputs, warranting a moderate circularity score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The feasibility claim rests on the CFD and CBCT simulation pipeline being faithful to clinical conditions and on the fully sampled FDK reconstruction being a correct ground truth. These are domain assumptions, not independently verified in the paper. The three patient geometries are treated as representative, which is an untested generalization assumption. No new physical entities are introduced.

free parameters (4)

CNN architecture hyperparameters = 64 initial filters, 8x8/5x5/3x3 kernels, learning rate 0.0001
Chosen without ablation; reconstruction quality depends on these settings, and no sensitivity analysis is provided.
Inlet flow velocities = 0.25, 0.35, 0.45 m/s
Hand-selected simulation conditions; only three levels tested, and results are not stratified by velocity.
Contrast injection durations = 0.5, 1.0, 1.5, 2.0 s
Hand-selected; no analysis of how injection duration affects reconstruction accuracy.
Projection truncation levels = up to 49 projections (29 and 49 per summary)
No systematic sweep or per-level performance reported.

assumptions (5)

domain assumption CFD contrast transport accurately represents in-vivo contrast dynamics.
Invoked in Section 2.1, where labeled particles simulate contrast bolus transit. No validation against patient data is provided.
domain assumption The simulated CBCT acquisition, including a 220-degree arc and 27 fps, faithfully models clinical C-arm systems.
Stated in Section 2.1 without comparison to real system geometry or noise characteristics.
domain assumption Parker-weighted FDK from 108 projections is a valid ground truth for 3D-QA.
Used in Section 2.2 as the training target; FDK artifacts are inherited by the evaluation.
domain assumption Three patient-specific geometries are sufficient to demonstrate generalizable feasibility.
Implicit in Sections 2.1 and 3; no independent test set from unseen patients is used.
domain assumption Test data drawn from the same three patients is representative of unseen anatomies.
The Results section reports only a pooled MSE, with no per-patient or leave-one-patient-out analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Leveraging Convolutional Neural Networks for 3D Quantitative Angiography Reconstructions from Sparse Cone Beam CT Projections Utilizing CFD Data." pith.science (2026). https://pith.science/paper/KYYIKPDC

@misc{pith2026241109632,
  author       = {Pith},
  title        = {Pith review of: Leveraging Convolutional Neural Networks for 3D Quantitative Angiography Reconstructions from Sparse Cone Beam CT Projections Utilizing CFD Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KYYIKPDC}},
  note         = {Machine review of arXiv:2411.09632}
}

read the original abstract

This study leverages convolutional neural networks to enhance the temporal resolution of 3D angiography in intracranial aneurysms focusing on the reconstruction of volumetric contrast data from sparse and limited projections. Three patient-specific IA geometries were segmented and converted into stereolithography files to facilitate computational fluid dynamics simulations. These simulations first modeled blood flow under steady conditions with varying inlet velocities: 0.25 m/s, 0.35 m/s, and 0.45 m/s. Subsequently, 3D angiograms were simulated by labeling inlet particles to represent contrast bolus injections over durations of 0.5s, 1.0s, 1.5s, and 2.0s. The angiographic simulations were then used within a simulated cone beam C arm CT system to generate in-silico rotational DSAs, capturing projections every 10 ms over a 220-degree arc at 27 frames per second. From these simulations, both fully sampled (108 projections) and truncated projection datasets were generated the latter using a maximum of 49 projections. High fidelity volumetric images were reconstructed using a Parker weighted Feldkamp Davis Kress algorithm. A modified U Net CNN was subsequently trained on these datasets to reconstruct 3D angiographic volumes from the truncated projections. The network incorporated multiple convolutional layers with ReLU activations and Max pooling, complemented by upsampling and concatenation to preserve spatial detail. Model performance was evaluated using mean squared error (MSE). Evaluating our U net model across the test set yielded a MSE of 0.0001, indicating good agreement with ground truth reconstructions and demonstrating acceptable capabilities in capturing relevant transient angiographic features. This study confirms the feasibility of using CNNs for reconstructing 3D angiographic images from truncated projections.

Figures

Figures reproduced from arXiv: 2411.09632 by the authors.

**Figure 2.** U-Net architecture. The network consists of a contracting path (left) for feature extraction and a symmetric expanding path (right) for precise localization. The contracting path downsamples the input image using convolutions and max pooling, while the expanding path upsamples the feature maps and combines them with corresponding features from the contracting path. 3. RESULTS U-Net model effectively reconstructed fu… view at source ↗

**Figure 3.** A) Phantoms generated by computational fluid dynamics from different angle projections, a) 0 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Reconstructed phantom images with varying contrast timing. The contrast agent was introduced into the vessel, and i [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 5 canonical work pages

[1]

Cerebral aneurysms and variations in the circle of Willis,

K. N. Kayembe, M. Sasahara, and F. Hazama, “Cerebral aneurysms and variations in the circle of Willis,” Stroke, 15(5), 846-50 (1984)

work page 1984
[2]

Color -coded digital subtraction angiography: the end of a monochromatic era?,

G. Benndorf, “Color -coded digital subtraction angiography: the end of a monochromatic era?,” AJNR Am J Neuroradiol, 31(5), 925-7 (2010)

work page 2010
[3]

Digital subtraction angiography (DSA): a neurosurgical perspective,

P. Butler, “Digital subtraction angiography (DSA): a neurosurgical perspective,” Br J Neurosurg , 1(3), 323-33 (1987)

1987
[4]

Digital angiography: a perspective,

C. A. Mistretta, A. B. Crummy, and C. M. Strother, “Digital angiography: a perspective,” Radiology , 139(2), 273-6 (1981)

work page 1981
[5]

Feasibility study for use of angiographic parametric imaging and deep neural networks for intracranial aneurysm occlusion prediction,

M. M. Shiraz Bhurwani, M. Waqas, A. R. Podgorsak, K. A. Williams, J. M. Davies, K. Snyder, E. Levy, A. Siddiqui, and C. N. Ionita, “Feasibility study for use of angiographic parametric imaging and deep neural networks for intracranial aneurysm occlusion prediction,” J Neurointerv Surg, 12(7), 714-719 (2020)

2020
[6]

Feasibility study of deep neural networks to classify intracranial aneurysms using angiographic parametric imaging

M. M. Shiraz Bhurwani, A. R. Podgorsak, A. Ramesh Chandra, R. A. Rava, K. V. Snyder, E. I. Levy, J. M. Davies, A. H. Siddiqui, and C. N. Ionita, "Feasibility study of deep neural networks to classify intracranial aneurysms using angiographic parametric imaging." 10950

work page
[7]

2D versus 3D comparison of angiographic imaging biomarkers using computational fluid dynamics simulations of contrast injections

Allison Shields, M. M. S. B, Kyle Williams, Venkat Keshav Chivukula, Daniel R. Bednarek, Stephen Rudin, Ciprian Ionita, "2D versus 3D comparison of angiographic imaging biomarkers using computational fluid dynamics simulations of contrast injections ". 12463-89

work page

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Cerebral aneurysms and variations in the circle of Willis,

K. N. Kayembe, M. Sasahara, and F. Hazama, “Cerebral aneurysms and variations in the circle of Willis,” Stroke, 15(5), 846-50 (1984)

work page 1984

[2] [2]

Color -coded digital subtraction angiography: the end of a monochromatic era?,

G. Benndorf, “Color -coded digital subtraction angiography: the end of a monochromatic era?,” AJNR Am J Neuroradiol, 31(5), 925-7 (2010)

work page 2010

[3] [3]

Digital subtraction angiography (DSA): a neurosurgical perspective,

P. Butler, “Digital subtraction angiography (DSA): a neurosurgical perspective,” Br J Neurosurg , 1(3), 323-33 (1987)

1987

[4] [4]

Digital angiography: a perspective,

C. A. Mistretta, A. B. Crummy, and C. M. Strother, “Digital angiography: a perspective,” Radiology , 139(2), 273-6 (1981)

work page 1981

[5] [5]

Feasibility study for use of angiographic parametric imaging and deep neural networks for intracranial aneurysm occlusion prediction,

M. M. Shiraz Bhurwani, M. Waqas, A. R. Podgorsak, K. A. Williams, J. M. Davies, K. Snyder, E. Levy, A. Siddiqui, and C. N. Ionita, “Feasibility study for use of angiographic parametric imaging and deep neural networks for intracranial aneurysm occlusion prediction,” J Neurointerv Surg, 12(7), 714-719 (2020)

2020

[6] [6]

Feasibility study of deep neural networks to classify intracranial aneurysms using angiographic parametric imaging

M. M. Shiraz Bhurwani, A. R. Podgorsak, A. Ramesh Chandra, R. A. Rava, K. V. Snyder, E. I. Levy, J. M. Davies, A. H. Siddiqui, and C. N. Ionita, "Feasibility study of deep neural networks to classify intracranial aneurysms using angiographic parametric imaging." 10950

work page

[7] [7]

2D versus 3D comparison of angiographic imaging biomarkers using computational fluid dynamics simulations of contrast injections

Allison Shields, M. M. S. B, Kyle Williams, Venkat Keshav Chivukula, Daniel R. Bednarek, Stephen Rudin, Ciprian Ionita, "2D versus 3D comparison of angiographic imaging biomarkers using computational fluid dynamics simulations of contrast injections ". 12463-89

work page