{"id":"5ea7a94c-e169-4de4-ba9e-3aabd3158ac4","arxiv_id":"2505.03463","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A backward-warping implicit neural representation with diffeomorphism regularization and motion-compensated analytical detail injection reconstructs nonperiodic dynamic CT images from one-projection-per-frame scans.","lead":"This paper introduces BIRD, a self-supervised method that reconstructs moving CT images from scans where each projection is taken at a different moment, such as one-beat cardiac imaging. It combines backward-warping neural fields, deformation smoothing, and analytical detail injection to reduce motion artifacts in extreme limited-angle settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline accuracy claim is not tested against the prior INR dynamic-CT methods it claims to surpass; only FDK, PICCS, and self-ablation variants are quantified.","rationale":"I read the method as a coherent integration of backward-warping INR, bidirectional GradICON-style diffeomorphism regularization, and PAR-based analytical detail injection. The XCAT simulations with ground truth, and the phantom/patient demonstrations, are real evidence that the pipeline can reconstruct moving objects under extreme limited-angle conditions. My concern is not with the internal mechanics but with the scope of comparison. The reader's weakest_assumption was identifiability of the topology-preserving/free-form decomposition; that is a legitimate theoretical gap, but the XCAT results against known ground truth already provide some behavioral evidence that the chosen prior works, so I would not make identifiability the deciding issue. The missing head-to-head INR comparison is more decisive because the paper's own novelty and conclusion are framed comparatively. Since this is exactly the kind of evidence that can be supplied without changing the method, the appropriate verdict remains conditional: accept the architectural contribution tentatively, but do not accept the \"more accurate than existing INR methods\" claim until the comparison is run. I did not find an internal inconsistency sufficient to reject the paper.","tokens_in":19232,"tokens_out":10073,"duration_ms":106104,"concrete_test":"Run PMF-STINR and STINR (using released implementations or a close reimplementation) on the same simulated XCAT 540-view projections at 120 bpm for both 0.25 s/rot and 0.5 s/rot, and on the 2500-frame clinical-cardiac simulation, with identical geometry, noise settings, and the same 10/20-phase ground-truth evaluation. Report per-phase PSNR/SSIM for FDK, PICCS, INR-base, INR-DM, INR-DM-AR, STINR, and PMF-STINR, including variance over random seeds. If INR-DM-AR does not beat the best prior INR method by a paired, statistically meaningful margin (e.g., Wilcoxon signed-rank over phases, or consistently >1 dB), the comparative headline is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is explicitly comparative: the abstract promises \"more accurate dynamic CT reconstruction\" and Section IV opens with \"superior reconstruction quality compared to existing techniques.\" The only quantitative baselines reported in Sections III-A and III-B are FDK, PICCS, and three self-ablation variants of BIRD (INR-base, INR-DM, INR-DM-AR). The two nearest INR-based dynamic CT methods, STINR [ref 34] and PMF-STINR [ref 35], are cited in Section I with specific alleged weaknesses (forward-warping cost, DVF smoothness/detail tradeoff), but they are never run on the same XCAT or clinical-cardiac data. Because those weaknesses motivate the entire design, the experiments as reported cannot distinguish \"BIRD's components genuinely fix prior INR failures\" from \"BIRD is a well-tuned INR that would look similar to or worse than prior INR methods under a fair comparison.\" This is not an internal inconsistency; it is that the main comparative assertion is unsubstantiated by the provided evidence. If an existing prior-model-free INR method matched INR-DM-AR on the identical XCAT projections and metrics, the abstract's accuracy claim would be false as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BIRD, a self-supervised implicit neural representation (INR) framework for nonperiodic dynamic CT reconstruction. The method models a dynamic object as the sum of a topology-preserving feature obtained by backward-warping a static image with a learned deformation vector field (DVF), and a free-form feature over the 4D coordinates. A bidirectional DVF module applies a diffeomorphism regularization loss adapted from GradICON, and a motion-compensated analytical reconstruction module generates a partial-angle-reconstruction (PAR) image that is fed into the free-form image prediction to enhance fine details. The framework is evaluated on simulated XCAT phantom data, simulated clinical cardiac data, a physical heart/lung phantom, and retrospective patient data, comparing against FDK, PICCS, and three self-ablation variants (INR-base, INR-DM, INR-DM-AR). The paper claims more accurate dynamic CT reconstruction than existing techniques, particularly for one-beat cardiac imaging and motion artifact removal in conventional scans.","tokens_in":19538,"tokens_out":3403,"duration_ms":35185,"significance":"If the claims are substantiated, BIRD addresses a clinically relevant problem: reconstructing high-resolution 4D images from one projection per temporal frame under nonperiodic motion, without patient-specific pre-scans. The framework has several commendable design elements: backward-warping reduces per-voxel computation compared to forward-warping; the diffeomorphism regularization is a principled way to constrain DVF plausibility; and the analytical reconstruction module attempts to recover high-frequency details that pure INR optimization tends to over-smooth. The paper also includes real projection data from commercial CT systems, which is valuable. However, the central comparative claim is currently not supported by the experiments: the closest prior INR-based dynamic CT methods criticized in the introduction are never included as baselines. The registration loss and analytical reconstruction module also raise circularity concerns that are not addressed. As a result, the work is promising but requires additional validation and analysis before its accuracy claims can be accepted.","major_comments":[{"comment":"The number of PAR groups K is a central hyperparameter in the analytical reconstruction module (Eqs. (20)-(21)), but it is never specified in the implementation details or in any experimental description. Likewise, the PAR reconstruction kernel and angular width of each projection group are not reported. Since the ablation INR-DM-AR versus INR-DM is used to justify the analytical module, the results are not reproducible without these settings. Please state K and the PAR reconstruction parameters for each experiment, and provide a sensitivity analysis or at least a short study of K's effect on PSNR and artifact level.","section":"Section II-D; Section II-B.3"}],"minor_comments":[{"comment":"The total loss in Eq. (26) uses λDM = 1.0 and λRGT = 0.1 as 'recommended for first attempt,' but the paper does not report how sensitive the results are to these weights. A short sensitivity study or a statement that the results are robust over a reasonable range would strengthen the paper.","section":"Section II-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope of the journal and addresses a timely problem. The main concern is the absence of the closest INR-based baselines, which is fixable but requires additional experiments. The circularity in Eq. (19) and the lack of identifiability analysis are also substantive but can be addressed with additional experiments and discussion. I do not see a fundamental flaw that would require rejection, provided the authors either add the missing comparisons or revise the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth engaging with, but the headline claim is not supported by the experiments as reported. BIRD combines three known ideas—backward-warping DVF, GradICON-style diffeomorphism regularization, and PAR-based analytical detail injection—into a single self-supervised INR framework for nonperiodic dynamic CT. That combination is new as far as I can tell from the cited prior work, and the backward-warping point is genuinely useful: direct per-voxel mapping avoids the whole-image DVF evaluation that forward-warping requires. The diffeomorphism loss is sensible and the DVF visualizations suggest it does what it says. The PAR-based detail injection is a nice way to get high-frequency information without patient-specific pre-scans. The dimensional-reduction 4D encoding is a modest but reasonable engineering choice.\n\nThe experiments are consistent with the claims: PSNR gains over FDK and PICCS, qualitative improvement on XCAT and clinical cardiac simulations, and two practical studies on real phantom and patient data. The self-ablation variants show each component contributes. That is real evidence the method works on these datasets.\n\nThe soft spots are real and, for a comparative paper, load-bearing. The abstract and Section IV claim 'more accurate' and 'superior' reconstruction compared to existing techniques, but the only quantitative baselines are FDK, PICCS, and BIRD's own ablations. STINR and PMF-STINR are cited in the introduction with specific weaknesses that motivate the entire design, yet they are never run on the same data. That means the experiments cannot distinguish 'BIRD fixes prior INR failures' from 'BIRD is a well-tuned INR that might look similar or worse under fair comparison.' That is the central issue.\n\nThe second issue is reproducibility. No code, no data, no error bars from repeated runs, and K (number of PAR groups) is never specified. Loss weights and hashgrid settings are given, but K is load-bearing for the analytical reconstruction module. Without it, the experiments are hard to assess or reproduce.\n\nThe registration loss (Eq. 19) uses images produced by the same network, so there is a feedback loop, but I would not call the central claim circular. The reconstruction is still supervised by measured projection data. The identifiability assumption behind the static-plus-free-form decomposition is unproven, but that is common in this literature and not a fatal flaw by itself.\n\nWho is this for? Someone in CT reconstruction or INR-based imaging who wants to see a thoughtful integration with real clinical motivation. It deserves peer review, but the reviewer should demand direct comparison with STINR and PMF-STINR, code/data release, and specification of K before the accuracy claim is accepted.","headline":"BIRD is a plausible, well-engineered INR framework for dynamic CT, but the paper's central accuracy claim is unsubstantiated because it never compares against the prior INR methods it criticizes.","tokens_in":20050,"tokens_out":2096,"would_cite":false,"duration_ms":21327,"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":"BIRD reconstructs nonperiodic cardiac CT from one projection per frame, removing motion artifacts without pre-scans.","keywords":["dynamic CT reconstruction","implicit neural representation","diffeomorphic deformation field","backward warping","nonperiodic motion","cardiac CT","motion artifact reduction","limited-angle reconstruction"],"falsifier":"A decisive test is to simulate a dynamic phantom with known ground-truth motion that includes an abrupt topological change—for example, a contrast agent splitting into two blobs—and known projections. If the free-form feature plus the diffeomorphic DVF cannot reproduce the split region while a method with a denser 4D prior can, the claimed representation capacity is insufficient; conversely, if BIRD reproduces the region but its DVF is wrong, the inverse-consistency loss alone is not enforcing true deformation.","tokens_in":19033,"feed_emoji":"🩻","tokens_out":7802,"duration_ms":74263,"temperature":0.7,"pith_summary":"BIRD is a reconstruction method aimed at CT scans of anatomy that moves nonperiodically, where every projection view is treated as its own instant in time. The paper's central claim is that a backward-warping implicit neural representation, kept approximately diffeomorphic, can turn such one-projection-per-frame data into a full four-dimensional image with sharply reduced motion artifacts. If correct, cardiac CT would no longer need to bin projections across several heartbeats, and routine scans spoiled by breathing or other motion could be reconstructed without a repeat acquisition or a patient-specific pre-scan. The paper shows on simulated phantoms, a moving heart-and-lung phantom, and a retrospective patient scan that the complete method beats conventional FDK and compressed-sensing baselines and that each added component improves detail and DVF quality.","feed_headline":"Nonperiodic cardiac CT rebuilt from one projection per frame","feed_subtitle":"BIRD pairs backward-warped neural fields with diffeomorphic motion estimates, cutting artifacts without any pre-scan.","key_machinery":"The load-bearing object is the backward-warping deformation model. A dynamic coordinate $(x_d,t)$ maps to a static coordinate $x_s = x_d + D_{d\\to s}(x_d,t)$, so one ray sample needs one query of the deformation network and one query of the static image network; forward-warping alternatives need the whole field. A paired forward field $D_{s\\to d}$ makes a cycle, and the method penalizes the Jacobian of the round-trip residual $x_s + D_{s\\to d}(x_s,t) + D_{d\\to s}(x_s + D_{s\\to d}(x_s,t), t) - x_s$, which keeps the deformation approximately diffeomorphic and anatomically plausible. A free-form feature encoded from the raw 4D coordinate extends the model to intensity or topological changes a warp cannot express. Motion-compensated partial-angle analytical reconstructions, aligned by the forward field and summed into a static space, feed high-frequency detail into the final image and modulate its texture through the reconstruction kernel.","core_discovery":"The central claim is that a nonperiodic dynamic attenuation field can be represented as two complementary features. The first is a topology-preserving feature: a static reference image queried at a static coordinate $x_s = x_d + D_{d\\to s}(x_d,t)$, so each dynamic voxel is produced by one backward-warping DVF query and one static-image query. The second is a free-form feature built directly from the 4D coordinate, which absorbs intensity and topological changes that a pure warp cannot represent. The paper further claims that pairing the backward DVF with a forward DVF and penalizing the Jacobian of their round-trip keeps deformations physiologically plausible, and that motion-compensated analytical reconstructions built from partial-angle images inject high-frequency detail into the final image without any pre-scan. In the experiments, the full method removes cardiac and respiratory motion artifacts that blur FDK and PICCS reconstructions, preserves small coronary vessels, and remains the best variant when the gantry rotation is slow enough that each cardiac phase is seen from an extreme limited-angle set.","pith_inferences":["Editorial inference: this dual-feature decomposition is a template for other one-shot dynamic inverse problems—4D cone-beam CT on C-arms, dynamic PET, or cine MRI—where each measurement is tied to a single instant.","Editorial inference: the paper leaves identifiability unproven; a simple stress test is to re-run BIRD from several random initializations on one projection dataset and measure whether the reconstructed 4D volumes agree away from the measured rays.","Editorial inference: the free-form feature is the least constrained part of the model, so a useful extension would be a minimal-explanation penalty that forces the DVF to carry as much of the motion as possible when topology is known to be preserved.","Editorial inference: because the analytical-reconstruction kernel can be adjusted to control final noise texture, a deployment-focused extension would be to select kernels to match a clinical site's preferred image appearance without retraining the network."],"forward_implications":["One-beat cardiac imaging becomes possible: a single cardiac cycle's worth of projections can yield a full 4D cardiac image without retrospective phase binning.","Conventional scans with nonperiodic motion (failed breath-hold, peristalsis, patient movement) can be reconstructed cleanly, potentially avoiding repeat scans.","Computational cost no longer scales with whole-volume DVF evaluation: each ray queries the backward DVF and static image exactly once.","No patient-specific historical CT is needed for detail preservation, because high-frequency information is recovered from the current scan's own projections through motion-compensated analytical reconstruction.","The diffeomorphism constraint gives a principled way to balance DVF complexity and anatomical plausibility in an ill-posed, one-projection-per-view setting."],"supporting_citations":[{"why":"Baseline method: prior image constrained compressed sensing that the paper compares against and outperforms.","marker":"[4]"},{"why":"Source of the partial-angle reconstruction motion-compensation idea used in the analytical enhancement module.","marker":"[9]"},{"why":"Existing forward-warping INR dynamic CT method whose computational and regularization limitations motivate BIRD.","marker":"[33]"},{"why":"Existing STINR method that relies on patient-specific prior DVFs and images; the no-pre-scan contrast.","marker":"[34]"},{"why":"Existing prior-model-free INR dynamic CT method; the main self-supervised INR competitor.","marker":"[35]"},{"why":"Multi-resolution hashgrid encoding used as the INR encoding block.","marker":"[36]"},{"why":"Source of the hyper-space free-form feature that extends the representation beyond topology-preserving warps.","marker":"[37]"},{"why":"Source of the approximate-diffeomorphism inverse-consistency Jacobian regularization.","marker":"[38]"},{"why":"Digital XCAT phantom used to generate ground-truth dynamic projection simulations.","marker":"[39]"},{"why":"VoxelMorph registration used to up-sample the 20-phase clinical cardiac images into the per-view frames of simulation study 2.","marker":"[40]"}],"fun_headline_variants":["BIRD: Backward-warped neural fields fix nonperiodic CT motion","One-projection CT: Diffeomorphic warp clears cardiac motion","Nonperiodic CT: Backward-warping INR beats motion artifacts","BIRD reconstructs cardiac CT from a single angle per beat","No pre-scan, no blur: BIRD's backward warp for moving CT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true 4D attenuation field can be decomposed into a smoothly warped static reference image plus a free-form feature, and that minimizing per-ray projection fidelity selects that decomposition uniquely; if the real motion cannot be expressed this way, or if several decompositions fit the data equally well, the reconstruction is not guaranteed to be the true image.","fun_headline_variants_meta":{"raw":{"variants":["BIRD: Backward-warped neural fields fix nonperiodic CT motion","One-projection CT: Diffeomorphic warp clears cardiac motion","Nonperiodic CT: Backward-warping INR beats motion artifacts","BIRD reconstructs cardiac CT from a single angle per beat","No pre-scan, no blur: BIRD's backward warp for moving CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3232,"prompt_tokens":1048,"completion_tokens":2184,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":2098}},"tokens_in":664,"tokens_out":2184,"duration_ms":14883,"temperature":1.0,"reasoning_tokens":2098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:51:13.023126+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to simulate a dynamic phantom with known ground-truth motion that includes an abrupt topological change—for example, a contrast agent splitting into two blobs—and known projections. If the free-form feature plus the diffeomorphic DVF cannot reproduce the split region while a method with a denser 4D prior can, the claimed representation capacity is insufficient; conversely, if BIRD reproduces the region but its DVF is wrong, the inverse-consistency loss alone is not enforcing true deformation.","supporting_citations":[{"cited_title":"Prior image constrained compressed sensing (PICCS): A method to accurately reconstruct dynamic CT images from highly undersampled projection data sets,","cited_arxiv_id":null,"evidence_quote":"Baseline method: prior image constrained compressed sensing that the paper compares against and outperforms."},{"cited_title":"Reduction of motion artifacts in cardiac CT based on partial angle reconstructions from short scan data,","cited_arxiv_id":null,"evidence_quote":"Source of the partial-angle reconstruction motion-compensation idea used in the analytical enhancement module."},{"cited_title":"Dynamic CT Reconstruction from Limited Views with Implicit Neural Representations and Parametric Motion Fields,","cited_arxiv_id":null,"evidence_quote":"Existing forward-warping INR dynamic CT method whose computational and regularization limitations motivate BIRD."},{"cited_title":"Dynamic cone-beam CT reconstruction using spatial and temporal implicit neural representation learning (STINR),","cited_arxiv_id":null,"evidence_quote":"Existing STINR method that relies on patient-specific prior DVFs and images; the no-pre-scan contrast."},{"cited_title":"Dynamic CBCT imaging using prior model-free spatiotemporal implicit neural representation (PMF-STINR),","cited_arxiv_id":null,"evidence_quote":"Existing prior-model-free INR dynamic CT method; the main self-supervised INR competitor."},{"cited_title":"Instant neural graphics primitives with a multiresolution hash encoding,","cited_arxiv_id":null,"evidence_quote":"Multi-resolution hashgrid encoding used as the INR encoding block."},{"cited_title":"HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields,","cited_arxiv_id":null,"evidence_quote":"Source of the hyper-space free-form feature that extends the representation beyond topology-preserving warps."},{"cited_title":"GradICON: Ap- proximate Diffeomorphisms via Gradient Inverse Consistency,","cited_arxiv_id":null,"evidence_quote":"Source of the approximate-diffeomorphism inverse-consistency Jacobian regularization."},{"cited_title":"4D XCAT phantom for multimodality imaging research,","cited_arxiv_id":null,"evidence_quote":"Digital XCAT phantom used to generate ground-truth dynamic projection simulations."},{"cited_title":"V oxelMorph: A Learning Framework for Deformable Medical Image Registration,","cited_arxiv_id":null,"evidence_quote":"VoxelMorph registration used to up-sample the 20-phase clinical cardiac images into the per-view frames of simulation study 2."}],"review_version":1}