{"id":"75c8e08e-a6ce-4e91-af53-99268a750ecb","arxiv_id":"2411.09632","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A modified U-Net reduced sparse-view artifacts in simulated 3D angiography volumes, but the evaluation uses FDK reconstructions as ground truth and lacks independent validation.","lead":"Researchers trained a U-Net to clean up incomplete 3D angiograms of simulated brain aneurysms using CFD-generated contrast data. The study is a proof-of-concept with only three synthetic patients and no clinical validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 1e-4 MSE is measured against FDK reconstructions, not the known CFD ground truth; if FDK deviates from true contrast concentration, the central feasibility claim is untested.","rationale":"The reader's weakest assumption is the validity of FDK ground truth, and I agree that this is the most load-bearing gap. The paper's only quantitative result is MSE = 1e-4, obtained by comparing the U-Net output to full-angle FDK reconstructions. If those FDK volumes are not faithful to the actual CFD contrast dynamics, the metric is uninformative. The synthetic setup makes the true volumes available, yet the paper never validates FDK against them, so the central feasibility claim is unsupported. This is not a disagreement with consensus but a correctness risk in the experimental logic: the measured quantity does not match the claimed quantity. The paper also lacks baselines, error bars, and public code/data, but those are secondary. My concern reinforces the reader's REJECT verdict; it does not shift it, so the verdict should remain UNCHANGED.","tokens_in":4163,"tokens_out":3065,"duration_ms":30149,"concrete_test":"Using the same synthetic pipeline, compute the full-angle (108-projection) FDK reconstruction for each CFD time step and compare it voxel-wise to the known CFD contrast concentration volume at that time step (the 'true' angiogram). Report MSE/SSIM and time-density curves. If the FDK-vs-CFD error is of the same order as or larger than the reported CNN-vs-FDK MSE of 1e-4, the 1e-4 metric does not establish reconstruction accuracy. As a stronger check, retrain the U-Net to map the truncated projection subsets directly to the CFD ground-truth volumes; if the resulting MSE is orders of magnitude worse, the central claim that the network reconstructs 3D angiography from sparse projections is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—CNN reconstruction of 3D angiographic volumes from truncated projections—rests on an unvalidated ground truth. In Section 2.2, training targets are 'the corresponding fully reconstructed CT slices,' i.e., Parker-weighted FDK volumes from 108 simulated projections over a 220-degree arc. 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. The network could achieve low MSE by learning to mimic FDK's specific artifact structure (cone-beam/reconstruction errors) rather than by recovering actual contrast dynamics. Because the CFD phantoms are synthetic, the true angiographic volumes are available, but the paper never compares either FDK reconstruction to them. Without this comparison, the feasibility claim—'acceptable capabilities in capturing relevant transient angiographic features'—is unsupported; the experiment demonstrates only that a U-Net can regress one reconstruction algorithm's output to another's. Additionally, the mapping is volume-to-volume (truncated FDK to full FDK), not projection-to-volume, so the abstract's phrasing overstates the setup. This makes the ground-truth validity the load-bearing assumption: if full-angle FDK is not an accurate proxy for true contrast dynamics, the quantitative result reported in Section 3 has no bearing on clinical 3D quantitative angiography.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":4443,"tokens_out":2756,"duration_ms":27626,"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":[{"comment":"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.","section":"Section 2.2 and Section 3"},{"comment":"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":"Abstract and Section 2.2"},{"comment":"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":"Section 3"},{"comment":"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.","section":"Section 3, Figure 4"}],"minor_comments":[{"comment":"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.","section":"Abstract/Summary"},{"comment":"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":"Figure 3"},{"comment":"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":"Section 2.1"},{"comment":"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.'","section":"Section 2.1"},{"comment":"Reference [6] is incomplete (no journal, volume, or pages) and reference [7] lacks the conference or proceedings title. Please complete the bibliographic entries.","section":"References"},{"comment":"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.","section":"General"}],"recommendation":"reject","confidential_remarks":"The core problem is that the evaluation does not test the stated claim. Because the ground truth is an FDK reconstruction from the same simulation pipeline, the experiment only demonstrates that a U-Net can map one FDK reconstruction to another. Adding a comparison against the known CFD contrast fields and a baseline MSE would be necessary to support feasibility, but such additions would require substantial new experiments. The paper might be better suited to a technical or conference venue after such revision, but in its current form the central claim is not established. There is also an apparent inconsistency between the abstract and the Summary regarding the number of projections (49 vs 29 and 49) that the editor may want the authors to resolve."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nOther 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.\n\nWho 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.\n\nWould 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.","headline":"Useful CFD-based simulation pipeline, but evaluating against FDK instead of the true contrast field leaves the feasibility claim untested.","tokens_in":5002,"tokens_out":2497,"would_cite":false,"duration_ms":24887,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["3D angiography","intracranial aneurysms","sparse projections","cone beam CT","computational fluid dynamics","convolutional neural networks","U-Net","quantitative angiography"],"falsifier":"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.","tokens_in":3982,"feed_emoji":"🧠","tokens_out":6652,"duration_ms":60819,"temperature":0.7,"pith_summary":"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.","feed_headline":"CNN rebuilds 3D angiograms from half the CT projections","feed_subtitle":"In simulated aneurysm scans, a U-Net matches full-angle reconstructions with an error of 1e-4.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the prior application of deep neural networks to intracranial-aneurysm angiography that motivates adding 3D information.","marker":"[5]"},{"why":"Companion prior work using deep networks with angiographic parametric imaging for aneurysm classification, further motivating the CNN approach.","marker":"[6]"},{"why":"Establishes the CFD-based simulated contrast-injection pipeline (2D versus 3D biomarkers) that this study extends to CNN reconstruction of truncated projections.","marker":"[7]"}],"fun_headline_variants":["AI rebuilds 3D angiograms from sparse CT projections","U-Net recovers full 3D blood flow from 49 projections","Sparse-scan CNN matches full 3D angiography in silico","Deep learning cuts CT projections for aneurysm imaging","Neural network predicts 3D contrast from partial scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["AI rebuilds 3D angiograms from sparse CT projections","U-Net recovers full 3D blood flow from 49 projections","Sparse-scan CNN matches full 3D angiography in silico","Deep learning cuts CT projections for aneurysm imaging","Neural network predicts 3D contrast from partial scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1482,"prompt_tokens":1029,"completion_tokens":453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":367}},"tokens_in":645,"tokens_out":453,"duration_ms":5109,"temperature":1.0,"reasoning_tokens":367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:25:51.940200+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Feasibility study of deep neural networks to classify intracranial aneurysms using angiographic parametric imaging","cited_arxiv_id":null,"evidence_quote":"Companion prior work using deep networks with angiographic parametric imaging for aneurysm classification, further motivating the CNN approach."},{"cited_title":"2D versus 3D comparison of angiographic imaging biomarkers using computational fluid dynamics simulations of contrast injections","cited_arxiv_id":null,"evidence_quote":"Establishes the CFD-based simulated contrast-injection pipeline (2D versus 3D biomarkers) that this study extends to CNN reconstruction of truncated projections."}],"review_version":1}