{"id":"e14c5e43-c417-4980-9f15-456b6ce1dc11","arxiv_id":"2506.22280","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DIGS couples 4D Gaussian Splatting with a low-rank free-form deformation model to reconstruct motion-compensated CBCT images per projection, claiming better image quality and 6x faster training than HexPlane.","lead":"Researchers introduce DIGS, a method for dynamic cone-beam CT reconstruction that models motion with a deformable Gaussian point cloud and B-spline free-form deformation. It reconstructs each projection as its own time point, avoiding phase sorting, and reports a 6x speedup over a HexPlane-based 4D Gaussian Splatting baseline on simulated data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation bias: cases 3-6 are generated with the same rank-2 B-spline FFD family the method assumes, so the reported image-quality gain over HexPlane may not generalize to higher-rank or non-B-spline motion.","rationale":"The reader's weakest-assumption analysis correctly identifies the load-bearing issue: the favorable results on cases 3-6 are generated with a motion model that matches the method's representational assumptions. My reading of the full text supports this. The method itself is internally consistent, and the ablation studies provide useful evidence that the FFD representation contributes to performance. However, the evaluation does not currently establish that the reported gains generalize to motions outside the rank-2 B-spline FFD family. Case 2 is the closest available probe, and its small gain (0.63 dB over HexPlane) suggests the benchmark match matters. No internal inconsistency or mathematical error was found in the DVF-to-Gaussian update derivation (Eqs. 7-8); the concern is about external validity of the evaluation, not the correctness of the derivation. The concrete test of K=5 PCA components would directly measure whether the method remains advantageous when the test motion has more degrees of freedom than the model. Since this is the same concern the reader raised and the conditional verdict already accounts for it, no change to the reader's verdict is needed.","tokens_in":8132,"tokens_out":5344,"duration_ms":63472,"concrete_test":"Create a new simulated dataset by repeating the Section 4.1 4DCT pipeline but retaining five PCA components (K=5) instead of two, then rerun Table 1 with the released code, keeping Nr=2 and HexPlane's settings identical to the paper. If the mean PSNR advantage of DIGS over HexPlane on this dataset falls below about 1 dB (the level seen in case 2), the claimed 'superior image quality' is an artifact of the matched low-rank B-spline test motion; if the advantage remains at least 2 dB, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—superior image quality over HexPlane at 6x speed—rests on six simulated datasets. For cases 3-6 (Section 4.1), ground-truth motion is built by registering 4DCT phases with NiftyReg (a cubic B-spline FFD) and then truncating the deformation stack to the mean plus two PCA components. The proposed model is exactly D(x,t)=x+sum_{r=1}^{2} omega_r(t) u_r(x), with each u_r a B-spline FFD (Eqs. 9-10) and Nr=2 (Section 4.2). Thus on 4 of 6 test cases the generating motion lies in (or very close to) the method's own hypothesis class. This is not circular—HexPlane can also fit the data—but the benchmark favors DIGS's inductive bias and likely inflates the reported PSNR advantage. The non-matched XCAT cases (1-2), especially case 2 with only 0.63 dB improvement over HexPlane, preview the performance gap when motion is not generated by the same low-rank B-spline model. The speedup is plausible because DIGS uses few parameters, but the 'sharper images/fewer artifacts' claim is not established for realistic motion of higher rank. A secondary concern is that HexPlane's run time is reported without stating whether its iterations or hyperparameters were matched to DIGS, so the 6x speedup may partly reflect optimization settings rather than representation efficiency alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DIGS, a 4D Gaussian-splatting method for dynamic CBCT reconstruction. Motion is represented explicitly by a low-rank free-form deformation model D(x,t)=x+∑_{r=1}^{Nr} ω_r(t)u_r(x), with cubic-B-spline spatial basis functions u_r and 1D B-spline temporal weights ω_r. Gaussian means are moved by D and covariances are updated by the Jacobian of D (Eqs. 7–8), so all Gaussian attributes evolve consistently under one DVF. The method jointly optimizes reference Gaussians and motion basis against measured projections. It is evaluated on six simulated CBCT datasets (two XCAT-based, four based on 4DCT patient data with PCA-compressed deformations) and compared with SuPReMo and a HexPlane-based 4DGS baseline, reporting better PSNR/RMSE and about a 6x speedup over HexPlane. Two ablation studies isolate the effect of the deformation-informed update and of the FFD spatial regularization. Code is provided.","tokens_in":8433,"tokens_out":3916,"duration_ms":45989,"significance":"If the reported gains are robust, DIGS would be a useful step toward practical dynamic CBCT: it is faster than a HexPlane-style 4DGS baseline, produces a continuous DVF that can be used for tumor tracking, and ships with code. The covariance update in Eq. (8) is standard fluid/continuum mechanics and is cleanly derived, and the ablation isolating the FFD spatial regularization is informative. However, the central performance claim is weakened by the evaluation design: four of the six test cases use ground-truth motion generated from exactly the same low-rank PCA B-spline FFD family that the method assumes (Nr=2, B-spline spatial basis). The non-matched XCAT cases show much smaller gains over HexPlane (e.g., 0.63 dB in case 2), so the generality of the claimed improvement is not established. The speedup is plausible but the comparison protocol with HexPlane is incompletely specified.","major_comments":[{"comment":"Cases 3–6 are generated by registering 4DCT phases with NiftyReg (a cubic B-spline FFD) and then applying PCA that retains only the mean deformation plus two principal components (§4.1), while the proposed model sets Nr=2 and uses B-spline FFD spatial basis functions (§4.2, Eq. 9–10). The generating motion therefore lies in the same hypothesis class as the method itself. The reported PSNR gains over HexPlane on cases 3–6 (roughly 2.1–2.7 dB) partly reflect a matched inductive bias rather than general superiority. The XCAT cases 1–2, which are not generated from this model family, show much smaller gains (case 2: 21.78 vs 21.15 dB). To support the central claim, please add evaluations with mismatched motion, e.g., retaining more than two PCA components (Nr=3 or 4), using non-B-spline deformations, or using independently measured respiratory motion, and report per-case results. Also report sensitivity to Nr.","section":"§4.1 and §4.2"},{"comment":"The claimed 6× speedup over HexPlane is based on wall-clock times of about 1h27m–1h50m for HexPlane versus 14–16m for DIGS, but the paper does not state whether the two methods were given the same number of training iterations, the same learning-rate schedule, or the same early-stopping criterion. Without this information or a runtime–quality curve, the speedup may partly reflect optimization settings rather than representation efficiency. Please specify the exact comparison protocol (iterations, convergence criteria, hardware, measured times) or provide performance-versus-time plots.","section":"§5, Table 1"},{"comment":"The conclusion that the method \"produces sharper images with fewer motion artifacts\" is supported primarily by mean PSNR/RMSE on the matched simulation cases and by a few example slices (Fig. 2). Since the ground truth is created by warping reference images with the simulated DVFs and the method explicitly models a DVF of the same family, the evaluation metric is aligned with the method's inductive bias. To strengthen the claim, please include quantitative artifact metrics or a reader study on the non-matched cases, and ideally on an independently acquired dataset or on clinical CBCT projections.","section":"§5, Discussion"}],"minor_comments":[{"comment":"The tilde notation is used for both the projected 2D Gaussian and the projected coordinates/covariance, which is confusing; consider using a distinct symbol, e.g., a subscript '2D'.","section":"§3.1, Eq. (4)"},{"comment":"Nr is set to 2 without any ablation or justification; since Nr is a key hyperparameter of the low-rank motion model, a sensitivity study (Nr=1,3,4) would help the reader understand its role.","section":"§4.2"},{"comment":"The description of the 4DCT-based simulation states that 310 time points are uniformly sampled from 20 breathing cycles, but it is not clear whether the projections are presented to the method in chronological order or shuffled; please clarify, since the temporal B-spline basis assumes a temporal ordering.","section":"§4.1"},{"comment":"Table 1 reports mean PSNR/RMSE over cases but not standard deviations or per-case values for the ablations; reporting these would help assess variability across datasets.","section":"§5, Table 1"},{"comment":"The claim that this is the first application of a physics-based deformation-informed framework to image reconstruction is strong given that PhysGaussian and related methods apply the same covariance update to Gaussian kernels; please clarify the distinction (image reconstruction versus graphics rendering) and soften the novelty claim if appropriate.","section":"§2.2"},{"comment":"Reference [9] is cited as an arXiv preprint; if a peer-reviewed version has appeared, please cite that version instead.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-bias concern raised in the main report is real and load-bearing: on 4 of 6 datasets the ground-truth motion is generated from the same rank-2 B-spline FFD family the method assumes, so the headline PSNR gains over HexPlane are partly a matched-model effect. The fix is straightforward—add mismatched or higher-rank motion tests and report the HexPlane comparison protocol—so I see this as a major-revision issue rather than a rejection. The method itself is clearly presented and the code availability is a plus. I would also ask the authors to moderate the 'first' claim in Section 2.2, since the deference to PhysGaussian and related works makes the novelty statement narrower than the abstract implies."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core method is clean: it takes the physics-based deformation idea from PhysGaussian, applies it to Gaussian mean/scale/rotation consistently via a Jacobian, and couples it with an explicit low-rank B-spline FFD motion model. That combination is genuinely new for CBCT reconstruction, and the implementation looks careful. Second, the headline claims rest on an evaluation that is partly rigged by construction. Cases 3–6 are generated by registering 4DCT phases with NiftyReg and then keeping the mean plus two PCA components, exactly the rank-2 B-spline FFD family the model assumes (Nr=2, B-spline spatial basis). This is not circular in the strict sense—HexPlane can also fit the data—but it hands DIGS a large inductive-bias advantage. Case 2, the one XCAT case not generated from this matched model, shows only a 0.63 dB PSNR gain over HexPlane, which is a big red flag for how the gains will look on real, higher-rank motion.\n\nWhat the paper does well: the derivation of Eq. 8 is standard and correct; the ablations are informative, especially Ablation Study 2 showing that the spatial regularization from FFD matters; the speedup from 1h40m to ~15min is plausible given the small number of parameters; and the code is promised. I also appreciate that the authors are honest about the trade-off in Ablation Study 1, admitting the decoupled model fits slightly better but the consistent DVF is the real advantage.\n\nSoft spots, in proportion: the evaluation bias is the main one, and the authors do not acknowledge it. They should either test on real patient data or simulate motion with more than two PCA components and non-B-spline deformations. Second, they claim the DI framework gives interpretable, consistent DVFs, but they never validate the estimated DVFs against the ground-truth DVFs. That is a missing experiment for a central claim. Third, the HexPlane runtime is reported without saying whether its iterations or hyperparameters were matched, so the 6x speedup may partly reflect optimization settings. These are fixable, and none of them sink the method's plausibility.\n\nBottom line: this is a solid methods paper for people working on 4D Gaussian splatting or motion-compensated CBCT. It deserves a serious referee, but it needs revision. I'd want the authors to add a DVF accuracy evaluation, broaden the motion model in the simulations, and either match HexPlane's settings or discuss the discrepancy. If they do that, I'd be happy to cite it. For now, I'd treat the quantitative gains as promising but unproven.","headline":"A coherent DI-4DGS method for dynamic CBCT with a plausible speedup, but the simulated evaluation is partly self-fulfilling because four of six test cases generate motion from the same rank-2 B-spline FFD family the method assumes.","tokens_in":8995,"tokens_out":1696,"would_cite":false,"duration_ms":20851,"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":"The paper claims a deformation-informed 4D Gaussian splatting method reconstructs dynamic CBCT about six times faster than a HexPlane baseline while producing sharper images.","keywords":["Dynamic CBCT","4D Gaussian Splatting","Free-Form Deformation","Low-rank motion model","Motion-compensated reconstruction","Deformation vector field","Cone-beam CT","Radiotherapy"],"falsifier":"Generate a test dataset whose ground-truth deformation is not low-rank B-spline, for example from a free-breathing trace with sliding organs or a five-component PCA motion model, and run the same comparison: if the PSNR gain over HexPlane falls to near the 0.6 dB seen in the unmatched XCAT case while the wall-clock advantage holds, the claim that the FFD model is the source of both speed and quality is weakened.","tokens_in":7889,"feed_emoji":"🩻","tokens_out":5100,"duration_ms":48166,"temperature":0.7,"pith_summary":"The paper takes aim at a clinical bottleneck: cone-beam CT scans for radiotherapy are corrupted by breathing motion, and the standard fix of sorting projections into respiratory phases breaks down when breaths vary. It proposes to reconstruct a full dynamic image sequence with 4D Gaussian Splatting, using an explicit low-rank free-form deformation model to move the Gaussian kernels instead of learning implicit neural motion. The central claim is that this deformation-informed design is both faster and sharper: about six times faster than the HexPlane-based 4DGS baseline, with higher PSNR and fewer motion artifacts across six simulated datasets. If the claim holds, dynamic CBCT becomes a practical, motion-resolved alternative to phase-binned reconstruction, and it produces a continuous deformation vector field that can be used for tumor tracking.","feed_headline":"Deformation-informed 4D Gaussian splatting cuts CBCT reconstruction time 6x","feed_subtitle":"Unified deformation field keeps every Gaussian consistent, yielding sharper dynamic CT images and a motion field for tumor tracking in…","key_machinery":"The load-bearing object is the deformation-informed free-form deformation model: a continuous DVF built as a low-rank sum of spatial B-spline basis functions (a control-point lattice with cubic B-spline interpolation) and temporal B-spline weights. Its job is twofold. It acts as the motion representation, so reconstruction estimates a small number of control-point displacements instead of per-Gaussian attributes, and it enforces consistency because every Gaussian attribute at time $t$ is derived from the same $\\boldsymbol{D}$ and its Jacobian $\\nabla_{\\boldsymbol{\\mu}}\\boldsymbol{D}$. The FFD also supplies spatial regularization: adjacent Gaussians are forced to move coherently, which the ablations show is the main source of the quality improvement.","core_discovery":"On its own terms, the discovery is that explicit, spatially regularized motion beats implicit neural motion for dynamic CBCT. The method, called DIGS, represents the attenuation volume as a cloud of 3D Gaussian kernels at a reference time and lets one continuous free-form deformation field $\\boldsymbol{D}(\\boldsymbol{x},t)=\\boldsymbol{x}+\\sum_{r=1}^{N_r}\\omega_r(t)\\boldsymbol{u}_r(\\boldsymbol{x})$ warp them: means move along the field and covariances transform by its Jacobian, so position, scale, and rotation stay mutually consistent. The spatial basis functions $\\boldsymbol{u}_r$ are interpolated from a learnable B-spline control-point lattice and the temporal weights $\\omega_r(t)$ from 1D B-splines, with $N_r=2$ in the experiments. This yields a $6\\times$ speedup over HexPlane (about 15 minutes versus 1 hour 40 minutes on the same GPU), higher mean PSNR (28.14 dB vs 25.86 dB), and a directly usable deformation vector field; ablations attribute the quality gain mostly to the FFD's spatial regularization.","pith_inferences":["The speed and quality advantages may shrink on real patient motion that is not well described by two PCA components; the XCAT cases in the paper already show a smaller gain (0.6 dB) when the ground truth does not come from the matched low-rank model.","The same deformation-informed coupling could be carried over to other tomographic reconstruction problems, such as PET or dynamic CT, where a per-projection motion field is needed.","Because the DI framework yields both images and DVFs, the estimated motion field could be checked against independent surrogates (e.g., breathing belts or MR-derived motion) to validate it on real scans.","A testable extension would be to increase $N_r$ adaptively during optimization, which could handle motion of higher spatial or temporal complexity without giving up the speed advantage."],"forward_implications":["Dynamic CBCT can be performed without phase sorting, so irregular breathing no longer has to be averaged into sorted phases.","A continuous deformation vector field is produced alongside the images, giving radiotherapy a voxel-wise motion estimate for tumor tracking and dose accumulation.","Reconstruction time drops to about 15 minutes on a single consumer GPU, roughly six times faster than the HexPlane 4DGS baseline, making the approach closer to clinical workflow.","Removing the FFD spatial regularization degrades mean PSNR by about 0.85 dB, showing that coherent Gaussian motion is what carries the quality gain."],"supporting_citations":[{"why":"Supplies the static 3D Gaussian splatting CBCT rendering and optimization backbone the method builds on.","marker":"[28]"},{"why":"Introduces the physics-based deformation of Gaussian covariances via the Jacobian that the deformation-informed update adopts.","marker":"[26]"},{"why":"The concurrent 4DGS CBCT method used as the HexPlane baseline and the source of the initial Gaussian sampling scheme.","marker":"[9]"},{"why":"The implicit motion representation that the paper compares against and claims to be six times slower.","marker":"[4]"},{"why":"Defines the free-form deformation and cubic B-spline interpolation that the spatial and temporal motion bases are built from.","marker":"[22]"},{"why":"Provides the patient 4DCT images whose deformation fields are PCA-reduced to create the matched test cases for cases 3-6.","marker":"[1]"}],"fun_headline_variants":["DIGS: 6x faster dynamic CBCT with deformation-consistent Gaussians","One deformation field for motion-consistent 4D Gaussian splatting in CBCT","Unified FFD motion speeds CBCT reconstruction 6x and boosts quality","Deformation-informed 4DGS: consistent motion, 6x speedup for dynamic CBCT","Low-rank FFD stabilizes 4D Gaussian motion for sharper dynamic CBCT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluations that show the largest improvement are generated with a motion model of exactly the same kind the method uses (two PCA components and B-spline free-form deformation), so the claimed advantage may rely on real motion being low-rank and smooth in the same way.","fun_headline_variants_meta":{"raw":{"variants":["DIGS: 6x faster dynamic CBCT with deformation-consistent Gaussians","One deformation field for motion-consistent 4D Gaussian splatting in CBCT","Unified FFD motion speeds CBCT reconstruction 6x and boosts quality","Deformation-informed 4DGS: consistent motion, 6x speedup for dynamic CBCT","Low-rank FFD stabilizes 4D Gaussian motion for sharper dynamic CBCT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1581,"prompt_tokens":1045,"completion_tokens":536,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":425}},"tokens_in":661,"tokens_out":536,"duration_ms":5753,"temperature":1.0,"reasoning_tokens":425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:06:50.964187+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a test dataset whose ground-truth deformation is not low-rank B-spline, for example from a free-breathing trace with sliding organs or a five-component PCA motion model, and run the same comparison: if the PSNR gain over HexPlane falls to near the 0.6 dB seen in the unmatched XCAT case while the wall-clock advantage holds, the claim that the FFD model is the source of both speed and quality is weakened.","supporting_citations":[{"cited_title":"In: Advances in Neural Information Processing Systems (NeurIPS) (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the static 3D Gaussian splatting CBCT rendering and optimization backbone the method builds on."},{"cited_title":"Spatiotemporal Gaussian Optimization for 4D Cone Beam CT Reconstruction from Sparse Projections","cited_arxiv_id":"2501.04140","evidence_quote":"The concurrent 4DGS CBCT method used as the HexPlane baseline and the source of the initial Gaussian sampling scheme."},{"cited_title":"International Journal of Radiation Oncology* Biology* Physics86(2), 372–379 (2013)","cited_arxiv_id":null,"evidence_quote":"Provides the patient 4DCT images whose deformation fields are PCA-reduced to create the matched test cases for cases 3-6."}],"review_version":1}