{"id":"a022c3d1-7fa5-4024-a687-e21677cb14b5","arxiv_id":"2501.04140","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A spatiotemporal Gaussian representation with a deformation network reconstructs 4D-CBCT from a 1-minute sparse-projection scan without prior images or explicit motion models.","lead":"This paper applies a 4D Gaussian splatting framework to reconstruct 4D cone-beam CT images from sparse X-ray projections, without using prior CT or motion models. It reports comparable image quality and lower target positioning error than prior-image-based methods on the AAPM SPARE benchmark, plus a demonstration on one clinical 1-minute scan.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PTV accuracy is only quantified on equi-spaced simulated projections, not on the clustered phase-binned geometry of a real 1-minute scan; the single clinical case lacks ground truth, so the central 1-minute-scan claim is not yet established.","rationale":"I read the paper as a method paper with the central claim that a 4D Gaussian representation with a deformation network can reconstruct 4D-CBCT from sparse projections without priors, and the strongest quantitative evidence is the PTV translation error on the SPARE validation cohort. I do not see an internal inconsistency or a fatal flaw; the method is coherent, the X-ray Gaussian formulation follows R2-Gaussian, and the code is released. However, the load-bearing gap is the mismatch between the quantitative evaluation setting (equi-spaced simulated projections with known phases) and the real 1-minute acquisition geometry (clustered angles, imperfect phase sorting) that motivates the paper. The authors acknowledge the mismatch and add a real clinical case, but without quantitative ground truth. Since the deformation network is not physically regularized, it could exploit the benign angular distribution of the SPARE data, and the Discussion's explicit admission that deformed Gaussians need not be realistic makes this plausible. A realistic-simulation test with the clinical acquisition pattern would settle this. This does not change the reader's CONDITIONAL verdict; it sharpens the condition: the method's 1-minute-scan claim should be validated with realistic angular clustering before being accepted as a clinical finding.","tokens_in":12884,"tokens_out":15525,"duration_ms":155672,"concrete_test":"Generate a realistic 1-minute acquisition from the SPARE ground-truth 4DCT by sampling projections at the exact angular positions and phase-bin assignments of the clinical Varian TrueBeam scan (half-fan, ~893 projections, 10 phases, clustered angles), adding phase-label jitter to mimic Amsterdam Shroud error, then reconstruct with the proposed method and compute PTV translation error against the known 4DCT ground truth using the SPARE evaluation code. If the PTV 3D error exceeds the equi-spaced SPARE value (1.65 mm) by more than, say, 1 mm, or if it no longer outperforms the six comparison methods, the central claim for real 1-minute scans is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the method reconstructs 4D-CBCT from a 1-minute scan with the lowest PTV translation error is supported quantitatively only on the AAPM SPARE simulated cohort. As the authors state in Section III.A, the SPARE clinical datasets generated with equi-spaced down-sampling 'do not accurately replicate the challenges of a real 1-minute CBCT scan'; in a true 1-minute acquisition, projections within each respiratory bin cluster around specific angles and phase sorting (e.g., Amsterdam Shroud) is imperfect. The authors therefore substitute a real 1-minute TrueBeam scan, but this single case is evaluated only qualitatively (Section III.E, Fig. 10), with no ground truth for PTV position. Thus the key quantitative claim—lowest PTV translation error—has not been demonstrated under the acquisition conditions that define the problem. The deformation network, which is not regularized by physics (Discussion: 'each deformed Gaussian does not necessarily represent realistic soft tissue deformation'), could, in principle, absorb streaking artifacts or phase-sorting errors into blob displacements, yielding low apparent error on equi-spaced data while degrading on clustered data. Without an experiment that combines realistic angular clustering and ground truth, this possibility is not excluded.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a 4D-CBCT reconstruction framework based on an X-ray-adapted 3D Gaussian splatting representation with a HexPlane deformation network. Canonical Gaussians initialized from an average CBCT are optimized against phase-binned measured projections, jointly with a network that deforms position, scale, rotation, and density as a function of phase index; voxelization yields the 4D volume. The method is evaluated on 29 simulated AAPM SPARE validation scans against FDK and five prior-based methods, reporting comparable RMSE/SSIM and the lowest PTV translation error among the compared methods, plus a single clinical TrueBeam 1-minute scan assessed qualitatively. The authors claim no prior CT or explicit motion model is needed and that the representation reduces unknowns by more than 100x compared with voxel grids.","tokens_in":13139,"tokens_out":8257,"duration_ms":74300,"significance":"If the quantitative claims hold, this would be a practically important step for 1-minute 4D-CBCT in radiotherapy: it replaces prior-image and explicit motion-model dependencies with a self-contained differentiable optimization, and the parameter-efficiency argument is appealing. The paper's strengths are its use of the public SPARE benchmark with organizer-provided evaluation code, comparison against six established methods, and public release of code and results. The evaluation is reproducible in structure. However, the current evidence does not yet establish the central clinical claim under real 1-minute acquisition geometry, and the reported performance margins are not supported by statistical tests. The method is nonetheless well grounded in prior Gaussian splatting and X-ray rasterization work, and the reported improvements are plausible.","major_comments":[{"comment":"The claim that the proposed method achieves the lowest PTV translation error and consistently better RMSE is not supported by any statistical significance analysis. The margins are small relative to the inter-patient variability visible in Figure 6 (e.g., 3D translational error 1.65 mm vs 1.79-1.90 mm for the closest competitors; rSI 0.93 deg vs 0.72 deg; AP 1.36 mm vs 0.99 mm). The paper reports RMS values over 29 cases only. Report paired differences with confidence intervals or p-values for the RMSE, SSIM, and PTV translation/rotation errors, and state which comparisons are meaningfully different.","section":"Section III.C, Table II, Figures 3-6"},{"comment":"The quantitative geometric accuracy of the method has not been demonstrated under the acquisition conditions that define the problem. The authors correctly note that the SPARE clinical datasets, generated with equi-spaced down-sampling, 'do not accurately replicate the challenges of a real 1-minute CBCT scan'; in a real 1-minute scan, projections within each respiratory bin cluster around specific angles and phase sorting is imperfect. The single real TrueBeam case is evaluated qualitatively only, without ground truth for PTV position. Therefore Table II does not establish the abstract's claim of reconstructing 4D-CBCT from a 1-minute scan with the lowest PTV translation error. Add a simulation study that combines realistic angular clustering (e.g., using the clinical scan's phase-bin angular distribution) with known ground truth, or a motion phantom with known target positions.","section":"Section III.A and III.E"},{"comment":"Because the deformation network is not regularized by physics, and the authors state that each deformed Gaussian 'does not necessarily represent realistic soft tissue deformation,' it remains possible that the network absorbs streaking artifacts or phase-sorting errors into blob displacements rather than true respiratory motion. This is directly relevant to the claimed contribution of 'accurately capturing underlying motion.' The paper provides no ablation or validation of the learned deformations. I recommend an ablation removing the density/scale/rotation deformation heads, and/or a comparison of predicted Gaussian displacements against known motion in a simulated phantom, to show that the motion-capture mechanism is real.","section":"Section IV (Discussion) and Section II.C"}],"minor_comments":[{"comment":"Equation (13): G{X', R', R', rho'} repeats R'; the second entry should be S' (the deformed scaling).","section":"Equation (13)"},{"comment":"Equation (12): The definition of f_h is difficult to parse; please clarify the index ranges for the resolution levels and the meaning of the product and union operations.","section":"Equation (12)"},{"comment":"The text says timestamps are 10 values (0 to 0.9), but Section III.E reconstructs 50 phases; specify whether the network is retrained with 50 phase bins or timestamps are rescaled.","section":"Section II.E and III.E"},{"comment":"The caption uses 'Phase 00' and 'Phase 50' while the clinical reconstruction is described as 10-phase; clarify the phase numbering.","section":"Figure 10 caption"},{"comment":"Use a distinct variable for the integration parameter (e.g., s) and give explicit integration limits to avoid the confusion of r(t) with the ray coordinate r.","section":"Equations (7)-(8)"},{"comment":"The figures in Section III.C appear to be ordered inconsistently: 'Fig. 6' is referenced before its caption and 'Fig. 5' appears in the caption sequence after that reference; renumber/reference the figures consistently.","section":"Section III.C figure ordering"}],"recommendation":"major_revision","confidential_remarks":"This is a competent engineering contribution with a strong public benchmark evaluation, but the central clinical claim needs the additional validation described in the major comments. I would support acceptance if the authors add a clustered-geometry simulation with ground truth and statistical significance tests; the current version is better framed as a feasibility study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague], quick take on arXiv:2501.04140.\n\nThis is a reasonably clean adaptation of two recent ideas—R2-Gaussian static CT reconstruction and HexPlane-based 4D Gaussian deformation—to 4D-CBCT from sparse projections. The combination is new for this application, and the authors evaluate it properly on the AAPM SPARE benchmark with code and results online. What they do well: no prior 4DCT, no explicit motion model, differentiable end-to-end training, and on 29 simulated validation cases they show image quality comparable to prior-image methods and the lowest PTV translation error among the six methods. The discussion is honest about limitations: 3-hour reconstruction time, no ablation, and the deformed Gaussians are not physics-regularized and may not represent real tissue deformation.\n\nWhere it goes soft, in proportion: the central clinical claim—geometric accuracy for real 1-minute scans—is not actually demonstrated. The quantitative PTV accuracy comes only from the SPARE simulated datasets, which use equi-spaced projection downsampling. The authors themselves state that this does not replicate real 1-minute 4D-CBCT, where projections cluster within respiratory bins and phase sorting is imperfect. The one real TrueBeam 1-minute case is evaluated only qualitatively (Fig. 10), no ground truth for PTV position, so the lowest-error claim hasn't been shown under the conditions the method is meant for. Also no significance tests on the RMSE/SSIM/PTV numbers, no ablation to show which deformation components matter, and there's a typo in Eq. (13) (R' listed twice). None of these kill the method's plausibility, but they mean the headline accuracy claim should be read as provisional.\n\nWho benefits: medical physics researchers working on 4D-CBCT and IGRT, and anyone comparing sparse-view reconstruction methods. It deserves a serious referee: the method is sound in concept, the benchmark is appropriate, and the code availability makes it useful. I'd send it to peer review with a request for statistical testing, an ablation, and—ideally—a validation that combines realistic angular clustering with ground truth, e.g., simulated clustered projections or a moving phantom. That missing experiment is the one that would make or break the 1-minute-scan claim.","headline":"A plausible, well-benchmarked combination of 4D Gaussian splatting and a deformation network for prior-free 4D-CBCT, but the headline accuracy claim rests on simulated equi-spaced projections, not real 1-min clustered sampling.","tokens_in":13697,"tokens_out":2851,"would_cite":true,"duration_ms":27959,"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 deforming cloud of Gaussian blobs reconstructs 4D cone-beam CT from a 1-minute scan without prior CT images.","keywords":["4D cone-beam CT","sparse-view reconstruction","Gaussian splatting","deformation network","image-guided radiotherapy","respiratory motion","X-ray rasterization","HexPlane"],"falsifier":"Reconstruct a digital phantom with known respiratory deformation from simulated sparse projections and compare the network's Gaussian displacements against the known ground-truth motion fields; if the projections match well but the displacement field does not reproduce the known motion, the geometric-accuracy result reflects artifact absorption rather than true motion recovery.","tokens_in":12687,"feed_emoji":"🩻","tokens_out":6664,"duration_ms":62092,"temperature":0.7,"pith_summary":"This paper claims that a 4D cone-beam CT of a breathing patient can be reconstructed from the sparse projections of an ordinary 1-minute acquisition by representing the patient as a deforming cloud of Gaussian blobs, without needing a prior CT or a prebuilt motion model. The authors show that jointly optimizing the blob properties and a neural deformation network against the measured X-ray projections removes streak artifacts while preserving respiratory motion. On a public sparse-view 4D-CBCT benchmark, the method matches the image-quality metrics of prior-image-based competitors and achieves the lowest 3D target translation error among six methods. A single clinical 1-minute scan was reconstructed into 10 and 50 respiratory phases, suggesting the approach could make 4D imaging practical for routine radiotherapy.","feed_headline":"Gaussian blobs give 4D lung CT from a 1-minute scan","feed_subtitle":"Joint blob-and-motion optimization matches prior-image methods and sets the lowest target shift error.","key_machinery":"The central object is the spatiotemporal Gaussian cloud: a set of roughly 80,000 to 350,000 3D Gaussian kernels, each described by an 11-parameter set (3D position, three scaling factors, four rotation parameters, and one density), rendered into 2D X-ray projections by density integration along rays. A deformation network, using HexPlane multi-resolution planes (the six 2D planes xy, xz, yz, xt, yt, and zt) and four MLP heads, maps a phase index to per-Gaussian changes in position, scale, rotation, and density, letting one canonical cloud represent all respiratory phases. Joint optimization of the Gaussians and the network minimizes the L1 plus SSIM mismatch between rendered and measured projections, while an adaptive control module densifies and prunes Gaussians during training. A CUDA-based voxelizer converts the deformed Gaussians into the final 4D-CBCT volumes.","core_discovery":"The paper's central discovery is that a 4D dynamic scene can be encoded as an explicitly deformable set of 3D Gaussians, each with a position, covariance, rotation, and density, and that optimizing those parameters directly against raw projection images yields phase-resolved CBCT volumes. A deformation network built on HexPlane spatiotemporal feature encoding feeds four small multilayer perceptrons that predict per-phase changes in position, scale, rotation, and density; the whole model is trained end-to-end with only L1 and structural-similarity projection losses. The final 4D volume is obtained by voxelizing the deformed Gaussians. The authors report that this parameterization, roughly 350,000 Gaussians and fewer than 100,000 network parameters, achieves comparable RMSE and SSIM to methods that use a prior four-dimensional CT and the lowest 3D translation error of the planning target volume among the six methods tested.","pith_inferences":["Editorial extension: Conditioning the deformation network on continuous time rather than discrete phase indices could remove phase sorting entirely; the authors list phase-sorting-free reconstruction as future work.","Editorial extension: The optimized Gaussian positions could act as an intrinsic tumor-motion surrogate, but the paper does not validate that individual Gaussians track the same tissue across phases, so this use would need independent verification.","Editorial extension: Because the central objective is projection consistency alone, a physics-based deformation regularizer would be the most direct stress test of whether blob displacements correspond to real tissue motion; the paper acknowledges such regularization is missing."],"forward_implications":["4D-CBCT no longer depends on a pretreatment CT or an explicit motion model, removing the main bias source in prior-image-based methods.","The compact parameterization (roughly 350k Gaussians plus under 100k network weights) replaces roughly 450 million voxel unknowns, cutting memory demands by two orders of magnitude and making high-resolution 4D reconstruction feasible on a single GPU.","On the public benchmark's validation cohort, the method gives the lowest 3D target translation error (1.65 mm) among six methods, supporting its use for target alignment before beam delivery.","A clinical 1-minute scan was reconstructed into both 10 and 50 respiratory phases, so temporal resolution is not capped by the acquisition time."],"supporting_citations":[{"why":"Supplies the public sparse-view benchmark, ground-truth 4DCT, comparison reconstructions, and evaluation code used for all quantitative results.","marker":"[2]"},{"why":"Provides the 3D Gaussian splatting scene representation, covariance decomposition, and adaptive densification and pruning that the method builds on.","marker":"[20]"},{"why":"Contributes X-ray rasterization, density voxelization, and the integration-bias correction that make Gaussian splatting valid for tomographic reconstruction.","marker":"[22]"},{"why":"Introduces the deformation-network-plus-3D-Gaussians design for 4D dynamic scenes that the paper adapts from natural-light rendering to X-ray imaging.","marker":"[23]"},{"why":"Supplies the HexPlane multi-resolution spatiotemporal encoding used to predict phase-dependent Gaussian deformations.","marker":"[24]"},{"why":"Provides the projection-based phase-sorting algorithm used to assign timestamps in the clinical 1-minute scan demonstration.","marker":"[25]"}],"fun_headline_variants":["Deformable Gaussians give 4D lung CT from sparse projections","Sparse-view 4D CBCT via spatiotemporal Gaussian optimization","One-minute 4D lung imaging with moving Gaussian blobs","Gaussian blobs beat prior-CT methods for 4D CBCT from sparse scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a fixed set of Gaussian blobs, deformed as a function of phase index, can represent true respiratory motion from projection consistency alone; the paper does not include an ablation or physics-based regularization that rules out the deformation network absorbing streaking artifacts and phase-sorting errors as blob displacements.","fun_headline_variants_meta":{"raw":{"variants":["Deformable Gaussians give 4D lung CT from sparse projections","Sparse-view 4D CBCT via spatiotemporal Gaussian optimization","One-minute 4D lung imaging with moving Gaussian blobs","Gaussian blobs beat prior-CT methods for 4D CBCT from sparse scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000821,"raw_usage":{"total_tokens":3639,"prompt_tokens":1038,"completion_tokens":2601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":2522}},"tokens_in":654,"tokens_out":2601,"duration_ms":18330,"temperature":1.0,"reasoning_tokens":2522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:39:47.133897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct a digital phantom with known respiratory deformation from simulated sparse projections and compare the network's Gaussian displacements against the known ground-truth motion fields; if the projections match well but the displacement field does not reproduce the known motion, the geometric-accuracy result reflects artifact absorption rather than true motion recovery.","supporting_citations":[{"cited_title":"SPARE: Sparse-view reconstruction challenge for 4D cone-beam CT from a 1 -min scan,","cited_arxiv_id":null,"evidence_quote":"Supplies the public sparse-view benchmark, ground-truth 4DCT, comparison reconstructions, and evaluation code used for all quantitative results."},{"cited_title":"3D Gaussian Splatting for Real-Time Radiance Field Rendering,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D Gaussian splatting scene representation, covariance decomposition, and adaptive densification and pruning that the method builds on."},{"cited_title":"R2 -Gaussian: Rectifying radiative Gaussian splatting for tomographic reconstruction,","cited_arxiv_id":null,"evidence_quote":"Contributes X-ray rasterization, density voxelization, and the integration-bias correction that make Gaussian splatting valid for tomographic reconstruction."},{"cited_title":"4D Gaussian Splatting for real -time dynamic scene rendering,","cited_arxiv_id":null,"evidence_quote":"Introduces the deformation-network-plus-3D-Gaussians design for 4D dynamic scenes that the paper adapts from natural-light rendering to X-ray imaging."},{"cited_title":"HexPlane: A Fast Representation for Dynamic Scenes,","cited_arxiv_id":null,"evidence_quote":"Supplies the HexPlane multi-resolution spatiotemporal encoding used to predict phase-dependent Gaussian deformations."},{"cited_title":"Robust breathing signal extraction from cone beam CT projections based on adaptive and global optimization techniques,","cited_arxiv_id":null,"evidence_quote":"Provides the projection-based phase-sorting algorithm used to assign timestamps in the clinical 1-minute scan demonstration."}],"review_version":1}