{"id":"1553f01e-fca5-4cbc-8308-ad147e6fc8dc","arxiv_id":"2412.12919","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"4DRGS models vessels as 4D radiative Gaussians with a neural attenuation field and reconstructs 3D vessel volumes from sparse-view dynamic DSA images in about 5 to 13 minutes, matching or beating prior state-of-the-art quality.","lead":"4DRGS reconstructs 3D blood vessel structure from a small set of X-ray DSA images in minutes instead of hours, using time-varying Gaussian kernels and fast rasterization. If confirmed, it could make 3D vessel imaging practical in the clinic with substantially less radiation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3D SOTA claim rests on CD/HD computed against an admittedly imperfect FDK reference meshed at a 3x different threshold than all reconstructions; the metric may measure attenuation calibration rather than geometry, and margins over VPAL are within one standard deviation.","rationale":"Reading the paper in good faith: the method is a sensible, well-engineered adaptation of R2-Gaussian to dynamic DSA, with a clean decomposition into time-invariant geometry and a compact time-varying attenuation field; the ablations support both innovations; and the runtime figures (12m38s versus VPAL's roughly 2h36m on the same RTX 3090) make the efficiency claim credible independently of any metric controversy. The paper also honestly lists its scope limits in Sec. 5 (no patient motion, no calibration errors, no single-/bi-plane DSA), which appropriately bounds the clinical claim to rotational neurovascular cases. The load-bearing weakness is the 3D evaluation protocol. The CD/HD numbers are the only quantitative evidence for the headline 'SOTA 3D reconstruction,' and that evidence depends on (a) a reference the authors themselves describe as inaccurate, (b) a threshold asymmetry whose direction and magnitude are uncontrolled because the volumes are on unknown attenuation scales, and (c) margins that are small relative to the reported standard deviations. Any one of these would be a reason to treat the 3D ranking cautiously; together they mean the 3D SOTA claim is unverified rather than demonstrated. My analysis agrees with the reader's weakest_assumption and sharpens it: the 3.1x threshold ratio corresponds to a physical iso-surface radius shift for Gaussian kernels, so the metric can differentially reward methods whose attenuation scale happens to sit near the chosen threshold. I also flag the ambiguity about which views feed the FDK initialization, since a full-view FDK init would undermine the sparse-view framing; this is secondary because the paper's phrasing and the R2-Gaussian precedent suggest sparse-view FDK. The 2D synthesis claim is stronger, because it is a genuine held-out-frame evaluation with consistent gains, but it validates consistency with the 2D projections rather than the correctness of the 3D volume. Overall the paper remains a strong systems contribution with a solid efficiency claim, but the central quality claim needs either corrected metrics, released code, or verification on public data before it can be taken as established. The reader's CONDITIONAL verdict already captures this; my concern confirms the condition is the right one, so no verdict change is needed.","tokens_in":21941,"tokens_out":12906,"duration_ms":121509,"concrete_test":"Recompute the Sec. 4.2 3D evaluation with scale-matched iso-surfaces: per case, estimate each volume's attenuation scale (e.g., the 99.5th percentile of non-zero voxels), normalize all algorithm volumes and the reference to a common scale, then mesh everything at one identical threshold (run both 0.008 and 0.025) with identical marching-cubes settings, and recompute CD/HD over all 15 cases with paired per-case Ours(30k), Ours(10k), and VPAL differences. If the ranking flips or any margin of at least 0.1 mm in CD or 0.5 mm in HD reverses, the 3D SOTA claim is not robust. While rerunning, log the projection views used for the FDK initialization in Sec. 3.1 to confirm that only the N training views are used, not the full 133-view set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central 3D-quality claim, SOTA or comparable CD/HD versus VPAL, rests entirely on a comparison against the Siemens inbuilt FDK volumes, which Sec. 4.1 itself admits are 'not entirely accurate.' Two uncontrolled choices make that metric hard to interpret. First, the meshing thresholds are asymmetric: 0.025 for the reference versus 0.008 for all algorithm volumes, a 3.1x ratio with no calibration justification, and the paper explicitly declines volume comparison 'due to the unknown data calibration issue.' For a Gaussian kernel with peak attenuation A and scale sigma, the iso-surface radius grows as sigma*sqrt(2*ln(A/theta)); a lower threshold therefore systematically thickens reconstructed vessels while the higher threshold thins the reference, and faint distal branches (small A) can vanish at either threshold depending on each volume's absolute attenuation scale. If the FDK-to-learned-volume scale offset varies across cases or methods, the CD/HD metric conflates attenuation calibration with geometry. Second, the reported margins are small relative to the variance: at 30 views CD is 1.72 +/- 0.29 versus VPAL 1.79 +/- 0.51, at 50 views VPAL wins CD (1.58 vs 1.67), and most PSNR/HD differences are within one standard deviation of a 15-case sample, so 'SOTA in most scenarios' is not statistically established even if the reference were perfect. Separately, Sec. 3.1 initializes kernels from 'a low-quality volume reconstructed by FDK' without stating which projection subset is used; if that FDK used the full 133-view set, the sparse-view comparison would be unfair, though the natural reading (sparse-view FDK, as in R2-Gaussian) is the likely implementation. The speedup claim (12-32x on the same RTX 3090) is robust and not in dispute, and the held-out 2D PSNR/SSIM gains survive, but 2D metrics do not by themselves certify 3D geometry.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces 4DRGS, a 4D radiative Gaussian splatting framework for reconstructing 3D vessel structures from sparse-view dynamic DSA images. Vessels are represented by Gaussian kernels with static geometry (position, rotation, scale) and a time-varying central attenuation predicted by a compact hash-encoded MLP (DNAF). The model is trained by differentiable X-ray rasterization against captured DSA images, with an accumulated-attenuation pruning criterion and a bounded scaling activation to improve quality. The final volume is obtained by voxelizing and averaging the trained kernels over time. Experiments on 15 real patient cases at 30, 40, 50, and 60 views compare against FDK, R2-Gaussian, TOGS, and VPAL using Chamfer/Hausdorff distances, PSNR/SSIM on held-out views, and training time. The paper claims state-of-the-art or comparable reconstruction quality in most scenarios with a 12x-32x speedup over VPAL. The 2D synthesis results are based on held-out views and appear internally consistent, but the 3D comparison relies on an admittedly inaccurate FDK reference meshed at a different attenuation threshold than the reconstructed volumes, which undermines the central 3D-quality claim as currently presented.","tokens_in":22293,"tokens_out":5078,"duration_ms":48825,"significance":"If the 3D-quality claim were substantiated, this would be a practically important contribution: it is the first Gaussian-splatting method for sparse-view dynamic DSA reconstruction, and the reported training times (about 13 minutes for the full version and 5 minutes for the fast version) are dramatically lower than the roughly 2.6 hours reported for VPAL. The core representation is well motivated: static vessel geometry with time-varying attenuation is a natural fit for DSA, and the two proposed components, accumulated attenuation pruning and bounded scaling activation, are clearly explained and ablated. The use of 15 real clinical cases and multiple view settings is a strength relative to many medical-imaging papers. However, the paper does not release code or data, and the main quantitative evidence for 3D superiority depends on a threshold-sensitive comparison against a reference that the paper itself describes as inaccurate. Because the 3D metric is load-bearing for the headline claim, the current evidence does not yet support the stated level of certainty.","major_comments":[{"comment":"The 3D evaluation against the FDK reference is not yet convincing. The paper states that \"the provided volumes are not entirely accurate\" and then uses them as reference geometry; it also meshes the reference at attenuation threshold 0.025 while meshing all reconstructed volumes at 0.008. Since the iso-surface of a thresholded Gaussian attenuation field depends strongly on the chosen threshold, a 3.1x asymmetry can systematically favor methods whose reconstructed attenuation scale is lower than the reference. CD and HD may therefore be measuring attenuation calibration rather than geometric accuracy. I ask for a threshold-sweep analysis (e.g., varying both thresholds and reporting CD/HD curves), a calibration procedure for attenuation scales across methods, or validation against a trustworthy ground-truth geometry (e.g., a phantom or a high-quality 3D-DSA volume) before the 3D superiority claim can be accepted.","section":"Sec. 4.1"},{"comment":"The claim of \"SOTA or comparable\" performance in most scenarios is not statistically supported by the reported numbers. At 30 views the Chamfer distance is 1.72 +/- 0.29 for 4DRGS versus 1.79 +/- 0.51 for VPAL; at 50 views VPAL wins CD (1.58 +/- 0.19 versus 1.67 +/- 0.29); and most PSNR and HD differences are within one standard deviation over 15 cases. The paper should report paired significance tests (e.g., Wilcoxon signed-rank) and effect sizes, and should identify explicitly at which view settings the differences are statistically significant. Without this, the abstract's \"SOTA\" phrasing overstates what the data show.","section":"Sec. 4.2, Table 2"},{"comment":"The method depends on several manually set quantities, including the FDK initialization threshold delta = 0.016, the scale bounds smin = 0.1 and smax = 10 times voxel spacing, and the pruning threshold epsilon = 1e-6, yet no sensitivity analysis is provided. Since the central claim is that the method works robustly for sparse-view clinical data, the paper should show how CD, HD, PSNR, and SSIM vary with these parameters over reasonable ranges, or otherwise justify that the reported results are not a tuned operating point.","section":"Sec. 3.1 and Sec. 3.2"},{"comment":"The efficiency comparison would be more complete if the paper reported convergence criteria and loss curves for VPAL and for 4DRGS. The current table reports final iteration counts for 4DRGS (10k and 30k) but does not state whether VPAL was run to its own convergence or was stopped at a fixed schedule. If VPAL was stopped early or was not tuned, the \"12x-32x speedup\" could partly reflect implementation choices. Reporting training loss over time for all methods would make the efficiency claim more robust.","section":"Sec. 4.1, competing methods"}],"minor_comments":[{"comment":"The row labeled \"Others\" is ambiguous; please list the cases it covers or give per-case configuration details in supplementary material.","section":"Sec. 4.1, Table 1"},{"comment":"The summation notation in Eq. (7) is typeset as \"P iter\" without a subscript; it should be a sum over iterations between neighboring pruning operations.","section":"Eq. (7)"},{"comment":"Please clarify whether the temporal perturbation tau is sampled once per image or per pixel/ray, since this affects the interpretation of the loss and the temporal-consistency claim.","section":"Eq. (6)"},{"comment":"In Eq. (3), s_i denotes the activated scale vector while earlier s_i is used as a scale parameter; the notation should be made consistent, for example by writing s_i as a vector and sigma as elementwise sigmoid.","section":"Sec. 3.1"},{"comment":"The TOGS volume baseline is reconstructed by rendering 720 views in a full circle and applying FDK; please justify the choice of a full circle rather than the original 198-degree arc, since this may affect the comparison.","section":"Sec. 4.1, TOGS baseline"},{"comment":"The paper does not state whether code or data will be made available; given the overlap of the authors with the VPAL and R2-Gaussian baselines, releasing source code or detailed baseline configurations would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"Two of the strongest baselines, VPAL and R2-Gaussian, share authors with this manuscript (VPAL: Z. Liu, H. Zhao, Z. Cui; R2-Gaussian: R. Zha and H. Li). This is not itself a reason to reject, but it raises the bar for ensuring that the baseline implementations and hyperparameters are not inadvertently disadvantaged. The larger concern is the evaluation protocol: the 3D quantitative claim rests on a threshold-asymmetric comparison against an admitted low-quality reference, and the headline \"state of the art\" is not supported by the reported standard deviations. I recommend major revision with a request for threshold sensitivity analysis, paired statistical tests, and a stronger reference or phantom validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a real new application — radiative Gaussian splatting moved from static CT to dynamic DSA — and the speedup is the headline number that holds up. The paper deserves a serious referee, but the 3D SOTA claim is conditional, not proven, because the metric is noisy and the reference has a known bias.\n\nWhat is actually new: the decomposition into time-invariant geometry plus a compact neural attenuation field over time is sensible and matches the physics of contrast flow. Accumulated attenuation pruning is a clean fix for the dynamic setting, where instantaneous attenuation would wrongly prune kernels that are merely unfilled at that timestamp. Bounded scaling activation is a small but useful correction for needle artifacts. The 12-32x speedup over VPAL on the same RTX 3090 is robust and not in dispute, and the held-out 2D PSNR/SSIM gains are consistent across view counts. That is meaningful engineering progress.\n\nWhere I would push back: the 3D comparison uses the inbuilt FDK volumes as references, although Sec. 4.1 admits they are \"not entirely accurate.\" The meshing thresholds are asymmetric — 0.025 for the reference, 0.008 for all algorithms — without a calibration argument. For a Gaussian kernel, the extracted iso-surface radius depends on the ratio between peak attenuation and threshold, so a lower threshold systematically thickens reconstructed vessels. If the learned volumes sit at a different absolute attenuation scale than the FDK reference, the CD/HD numbers measure calibration more than geometry. The margins over VPAL are also small relative to the spread: at 30 views CD is 1.72±0.29 vs 1.79±0.51, and at 50 views VPAL wins CD outright. A 15-case sample with overlapping error bars does not support \"SOTA in most scenarios\" as a statistical statement. The TOGS baseline, forced through 720-view rendering plus FDK, also likely understates what TOGS could do. None of this kills the contribution — the speed and the 2D synthesis are enough — but it means the 3D claim should be softened to \"comparable to VPAL\" until verified on public data or with released code.\n\nThis paper is for the medical imaging and 3D reconstruction crowd. I would send it to a serious referee with a request to fix the evaluation protocol, not desk reject it. The core idea is sound, the engineering is honest, and the limitations section candidly notes patient motion and the small private dataset.\n\nRecommendation: engage; require better 3D evaluation before accepting.","headline":"Genuinely new and fast Gaussian-splatting pipeline for dynamic DSA reconstruction, but the 3D SOTA claim hinges on an asymmetric-threshold metric with margins inside the noise.","tokens_in":22955,"tokens_out":4639,"would_cite":true,"duration_ms":37361,"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":"This paper claims that 3D vessel reconstruction from sparse dynamic DSA images can be done in minutes with 4D radiative Gaussian splatting, a decoupling of static vessel geometry from time-varying contrast attenuation, at quality matching…","keywords":["4D radiative Gaussian splatting","sparse-view DSA reconstruction","3D vessel reconstruction","dynamic digital subtraction angiography","contrast-agent flow modeling","Gaussian kernel pruning","X-ray rasterization","neural attenuation field"],"falsifier":"Run 4DRGS and the previous method on a synthetic or physical phantom with known vessel centerlines, generate sparse-view dynamic DSA projections from a known ground-truth attenuation volume, and measure Chamfer and Hausdorff distances against the true geometry at identical mesh thresholds; if the gap between methods shrinks or reverses, or if the 0.025-versus-0.008 threshold choice explains the reported margin, then the paper's accuracy claim is a measurement artifact.","tokens_in":21705,"feed_emoji":"🩻","tokens_out":6320,"duration_ms":52324,"temperature":0.7,"pith_summary":"This paper tries to establish that high-quality 3D vessel reconstruction from sparse-view dynamic digital subtraction angiography (DSA) images can be done in minutes rather than hours by modeling vessels as a set of 4D radiative Gaussian kernels. The key move is to split the problem: each kernel has fixed geometry (position, rotation, and scale) that encodes the static vessel structure, while a compact neural network predicts its time-dependent central attenuation to follow contrast-agent flow. The kernels are trained by differentiable X-ray rasterization against real DSA images, then voxelized into a 3D volume. On 15 patient cases with 30 to 60 input views, the method reaches state-of-the-art or comparable reconstruction and synthesis quality while training in about 13 minutes, and a fast 10k-iteration version matches the previous method's quality in about 5 minutes, a 32x speedup. If this holds, it would make 3D DSA reconstruction practical enough for clinical use with substantially lower radiation exposure.","feed_headline":"3D vessel maps from sparse DSA images in 5 minutes","feed_subtitle":"A 4D Gaussian-splatting model cuts reconstruction time from hours to minutes while preserving vessel detail.","key_machinery":"The load-bearing object is the 4D radiative Gaussian kernel $G_i(x,t) = \\rho(p_i,t)\\exp\\left(-\\frac{1}{2}(x-p_i)^\\top \\Sigma_i^{-1}(x-p_i)\\right)$, where $\\Sigma_i = R_i S_i S_i^\\top R_i^\\top$ encodes fixed orientation and size and $\\rho(p_i,t)$ is the time-varying central attenuation predicted by a compact neural field. This single identity does two jobs: it constrains the reconstruction to static vessel geometry, which regularizes the severely ill-posed sparse-view problem, and it lets contrast-agent dynamics be represented as a scalar function of position and time rather than as a full 4D volume. X-ray rasterization from radiative Gaussian splatting integrates these kernels along rays to synthesize DSA images, and GPU voxelization converts the trained kernels into an attenuation volume. Accumulated attenuation pruning and bounded scaling activation are the two supporting mechanisms that keep the kernel set clean and compact during optimization.","core_discovery":"The authors claim that 4DRGS is the first Gaussian-splatting framework for 3D vessel reconstruction from sparse-view dynamic DSA images. The central discovery is that time-varying contrast-agent flow and static vessel anatomy can be decoupled in a single explicit representation: vessels are static in space, so only attenuation needs to change over time. Each 4D radiative Gaussian kernel carries time-invariant position, rotation, and scale, and a neural attenuation field predicts the kernel's central attenuation from its position and timestamp. The kernels are optimized with an L1 plus D-SSIM loss against real DSA frames under temporal perturbation, and the trained kernels are voxelized and averaged across timestamps to produce the final vessel volume. The paper further introduces accumulated attenuation pruning, which removes kernels whose attenuation averaged over training is consistently small, and bounded scaling activation, which keeps kernels within a fixed size range to avoid needle artifacts. With these components, the method reports state-of-the-art or comparable Chamfer and Hausdorff distances for 3D reconstruction and PSNR/SSIM for 2D DSA synthesis across 30, 40, 50, and 60 input views, while training in minutes.","pith_inferences":["Editorial inference: the reported 3D accuracy is measured against scanner-inbuilt FDK volumes that the paper itself calls 'not entirely accurate,' and reference meshes are extracted at a higher attenuation threshold (0.025) than reconstructed meshes (0.008); an independent ground-truth phantom study would determine how much of the Chamfer/Hausdorff margin is real geometry fidelity.","Editorial inference: the static-geometry assumption means the approach would likely degrade with patient motion or cardiac-driven vessel displacement; extending the kernels with per-kernel temporal displacement or motion compensation is the natural next test.","Editorial inference: the same 'static structure, time-varying attenuation' decomposition should transfer to other dynamic tomographic problems, such as contrast-enhanced cone-beam CT or 4D CT, where a compact Gaussian-plus-MLP representation could replace full 4D volume optimization.","Editorial inference: because the neural attenuation field is compact and rasterization is differentiable, the trained 4DRGS model could be used for near-real-time re-rendering of DSA frames at arbitrary angles, which would help interventional navigation; the paper does not demonstrate this but the representation supports it."],"forward_implications":["A full 3D vessel reconstruction from 30 to 60 DSA views converges in about 13 minutes on a single GPU, and a fast 10k-iteration version reaches the previous method's quality in about 5 minutes, a 12x to 32x speedup.","The method reports the best or second-best Chamfer and Hausdorff distances against scanner-volume references and the best PSNR/SSIM on held-out DSA frames in most tested view counts.","Because kernels only cover vascular structures rather than the whole scan volume, training and rendering time scale with vessel sparsity instead of scene size.","Accumulated attenuation pruning preserves vessels that are not yet opacified at a given timestamp, which random or instantaneous-threshold pruning removes.","Bounded scaling activation removes the needle-like elongated Gaussians that unbounded exponential scaling produces in DSA reconstruction."],"supporting_citations":[{"why":"provides the radiative Gaussian kernel formulation and the differentiable X-ray rasterization and voxelization modules that the method builds on","marker":"[30]"},{"why":"supplies the Gaussian kernel parameterisation, covariance construction, and adaptive-control densification strategy","marker":"[11]"},{"why":"is the state-of-the-art NeRF-based DSA reconstruction baseline that the speedup and quality comparisons must beat","marker":"[15]"},{"why":"provides the FDK algorithm used to initialise kernel positions from a low-quality volume and as a competing reconstruction baseline","marker":"[8]"},{"why":"is the 3DGS-based DSA synthesis baseline whose random pruning the paper contrasts with accumulated attenuation pruning","marker":"[32]"},{"why":"supplies the multiresolution hash encoding used in the compact neural attenuation field","marker":"[18]"}],"fun_headline_variants":["3D vessels from sparse DSA in just 5 minutes","4D Gaussian splatting reconstructs vessels 32x faster","Vessel maps from sparse views in 5 minutes flat","Radiative Gaussian splatting for rapid 3D vessel imaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 3D accuracy rests on the scanner's inbuilt FDK volumes being accurate enough to serve as reference geometry, even though the paper says they are 'not entirely accurate,' and on comparing surfaces that were meshed at different attenuation thresholds for reference versus reconstruction.","fun_headline_variants_meta":{"raw":{"variants":["3D vessels from sparse DSA in just 5 minutes","4D Gaussian splatting reconstructs vessels 32x faster","Vessel maps from sparse views in 5 minutes flat","Radiative Gaussian splatting for rapid 3D vessel imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1497,"prompt_tokens":1012,"completion_tokens":485,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":412}},"tokens_in":628,"tokens_out":485,"duration_ms":4725,"temperature":1.0,"reasoning_tokens":412,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:35:32.775111+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run 4DRGS and the previous method on a synthetic or physical phantom with known vessel centerlines, generate sparse-view dynamic DSA projections from a known ground-truth attenuation volume, and measure Chamfer and Hausdorff distances against the true geometry at identical mesh thresholds; if the gap between methods shrinks or reverses, or if the 0.025-versus-0.008 threshold choice explains the reported margin, then the paper's accuracy claim is a measurement artifact.","supporting_citations":[{"cited_title":"Josa a 1(6), 612–619 (1984)","cited_arxiv_id":null,"evidence_quote":"provides the FDK algorithm used to initialise kernel positions from a low-quality volume and as a competing reconstruction baseline"}],"review_version":1}