{"id":"189b0737-e599-4cdf-85be-8bde786b7528","arxiv_id":"2506.10309","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A deep unrolling network with spatiotemporal rotation equivariance outperforms previous methods on cardiac cine MRI at high undersampling factors.","lead":"Dynamic MRI reconstruction is hard when only a fraction of the data is collected. This paper builds a deep unrolling network that is designed to be invariant under rotating images in space and across time, and reports sharper cardiac cine MRI reconstructions than comparison methods at acceleration factors from 8x to 24x.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global equivariance claim is not established: Eq. (15) assumes a rotation-equivariant learned DC update, but the Cartesian VISTA forward operator does not commute with 90° rotations, and the temporal layer's 'trivial group' makes temporal rotation equivariance nominal.","rationale":"I read the paper in good faith. The empirical core—consistent SOTA PSNR/SSIM/HFEN gains at R=8–24 on an in-house cine dataset, plus qualitative generalization on OCMR—is reported carefully and is not the object of this concern. The central scientific claim, however, is the 'rigorous propagation' of spatiotemporal rotation equivariance through the unrolled network. That claim requires every component, including the data-consistency step, to be exactly equivariant. The paper proves this only under an idealized unitary setting and then asserts it for the learned SREC-DC network in Eq. (15). In actual dynamic MRI with Cartesian VISTA undersampling, the measurement operator is not rotation-equivariant, so a DC update that uses the measured k-space cannot be exactly equivariant. This is a correctness risk to the central claim, not merely a disagreement with current consensus. The reader's weakest_assumption correctly identifies the temporal layer, and the reader also flags Eq. (15) in the rationale; my stress-test agrees with these observations but elevates the DC issue because it breaks equivariance even for spatial rotations, which is the stronger, more fundamental part of the claim. A conditional verdict remains appropriate if the authors can either prove equivariance under explicit conditions (e.g., rotation-symmetric masks) or reframe the claim as an equivariant prior with approximate DC; without such a clarification the symmetry-based novelty is overstated. The absence of released code/data lowers confidence in the empirical audit, so the proposed commutator and equivariance-error checks are the decisive next steps.","tokens_in":21049,"tokens_out":12920,"duration_ms":166921,"concrete_test":"Analytically compute the commutator [R90, A^H A] for A = PΩFS with the actual R=12 VISTA mask and ESPIRiT maps used in the paper; nonzero commutator entries at unsampled frequencies falsify the premise of Eq. (14) and hence Eq. (15). Complement this with a numerical equivariance audit on a trained DUN-SRE: rotate a test input slice, feed the same mask/coil sensitivities, and compare the output with the rotated output; report the relative equivariance error. If the authors intend Q_SREC to be a pure function of x, they must state this and explain how k-space measurements enter the update.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III-C2's Remark 1 proves equivariance of an explicit L2 DC term under a unitary ρ, but omits the required hypothesis that the forward operator A commutes with ρ. For the actual acquisition A = PΩFS (Cartesian VISTA undersampling, Fourier encoding, ESPIRiT sensitivities), a 90° rotation R satisfies R PΩ ≠ PΩ R because the mask is anisotropic. Consequently x − η A^H(Ax − y) is not rotation-equivariant, and the learned SREC-DC used in Eq. (15) inherits this unless it ignores y and the mask entirely, which is not stated and would sever data consistency. The temporal layer (Sec. III-B4) is separately declared to have SG act 'as a trivial group in practical implementation,' so its cyclic channel shifts propagate the spatial group index across time but do not introduce a genuine temporal rotation symmetry. Together these two gaps mean the abstract's claim of 'rigorous propagation of spatiotemporal rotation symmetry constraints throughout the reconstruction process' is unsupported; the reported SOTA can still stand, but cannot be attributed to exact spatiotemporal equivariance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DUN-SRE, a deep unrolling network for dynamic MRI reconstruction that aims to embed rotation equivariance in both the spatial and temporal dimensions via a (2+1)D equivariant convolution architecture. The network consists of a spatiotemporal rotation equivariant proximal module (SREC-Prox) and an equivariant data consistency module (SREC-DC), along with a Fourier-basis filter parametrization intended to preserve representation fidelity. The method is evaluated on an in-house cardiac cine dataset and the public OCMR dataset, reporting state-of-the-art PSNR, SSIM, and HFEN at acceleration factors R=8,12,16,20,24, with ablation studies on the proximal module, the DC module, and the filter parametrization.","tokens_in":21283,"tokens_out":7401,"duration_ms":84719,"significance":"If the equivariance claims were correct, DUN-SRE would be a notable contribution to symmetry-aware deep learning for dynamic imaging, potentially improving reconstruction quality at high acceleration by leveraging a physically motivated prior. The paper provides a substantial experimental comparison with multiple baselines, per-metric means and standard deviations, and a 3-fold cross-validation ablation, which are strengths. However, the central theoretical claim of rigorous spatiotemporal rotation equivariance is undermined by the trivial temporal group and the non-commuting forward operator in the data consistency term; the empirical gains may still stand, but they cannot be attributed to exact spatiotemporal equivariance as the paper states.","major_comments":[{"comment":"Remark 1 claims that an L2-norm data consistency update is equivariant to any unitary transformation ρ. This assertion is only valid if the forward operator A commutes with ρ and if the measured data y transforms accordingly under the group action. For the actual acquisition A = PΩFS with the Cartesian VISTA mask, the sampling mask PΩ is anisotropic and does not commute with 90° rotations: a mask that samples only vertical k-space lines, applied to a rotated image, still samples the same vertical lines rather than the rotated set. Consequently the gradient descent step x − η A^H(Ax − y) is not rotation-equivariant. The SREC-DC network, which takes the raw k-space data and ESPIRiT sensitivity maps as inputs, inherits this non-equivariance unless it entirely ignores y and the mask, which would sever data consistency. This invalidates the global equivariance assertion in Eq. (15).","section":"III-C2 (Remark 1)"},{"comment":"The temporal equivariant layer is defined with a group SG that is explicitly stated to act as a trivial group in practical implementation, composed only of the identity element (Section III-B4). A trivial group action does not implement any temporal rotation symmetry; the cyclic channel shifts only permute the spatial group index across time and do not rotate the temporal axis. Therefore the network enforces spatial rotation equivariance only, and the claims of 'spatiotemporal rotation equivariance' and 'temporal rotation symmetry constraints' in the abstract and introduction are not supported. The empirical gains attributed to temporal equivariance could instead arise from the (2+1)D factorization and weight sharing across frames. The authors should either implement a non-trivial temporal group action or revise the framing to channel-level group shuffling and temper the symmetry claims accordingly.","section":"III-B4"},{"comment":"The derivation of global equivariance is only a chain of equalities with no formal proof or numerical verification. Equation (15) is also not well-typed: applying Q_SREC to ρ_g and then to X^k is ambiguous, and the composition of operators is not clearly defined. More fundamentally, even if H_SREC and Q_SREC were each equivariant, the composition is equivariant only if the group action on every input to the DC network, including the measured k-space data y and the ESPIRiT sensitivity maps, is compatible with the image-space rotation. The manuscript does not define such an action, and for fixed y and maps the composition cannot be equivariant. We suggest explicitly stating the group action on all inputs and providing an equivariance error measurement (e.g., applying random 90° rotations to test inputs and measuring the output disagreement) to support the claim.","section":"III-C2 (Eq. (15))"}],"minor_comments":[{"comment":"At R=24, DL-ESPIRiT and U-MCMR report identical PSNR (38.6397) and SSIM (0.9373) values but different HFEN values; this appears to be a transcription or copying error and should be corrected.","section":"Table III"},{"comment":"There are several typos in the manuscript, including 'imtermediate' in Section III-B4, 'dicrete' in Section III-B1, 'denonted' in Section III-C1, and 'nework' in the Conclusion; these should be fixed.","section":"III-B4"},{"comment":"Equation (17) appears to have a missing closing parenthesis in the definition of the 1D Fourier basis filter, and the summation index and limits should be checked for consistency with the filter size p.","section":"III-D"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concerns are valid: the paper's central theoretical claim of exact spatiotemporal rotation equivariance is not supported by the present manuscript, but the empirical contribution may still be valuable. I recommend major revision rather than rejection because the authors could reframe the contributions as approximate or channel-level group-structured temporal feature sharing and add supporting equivariance-error experiments, while retaining the demonstrated reconstruction gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me cut to it. The paper's empirical core looks solid: DUN-SRE beats the listed baselines at all five acceleration rates on the in-house cine dataset, with means and standard deviations, and the ablation ladder (VCNN, 2D-ECNN, SREC-Prox, SREC-ProxDC, DUN-SRE) is coherent and shows each added piece helps. The generalization and prospective OCMR experiments are qualitative but not cherry-picked-looking. Whoever reads the tables will get a fair picture of the method as an architecture.\n\nWhat's actually new is the (2+1)D equivariant unrolling: 2D group-equivariant layers for spatial rotation, a 1D temporal equivariant layer that cyclically shifts group channels to keep the group structure, and both proximal and DC modules built from the same equivariant blocks. That combination is not in the cited prior work, including the authors' own SRE-CNN. The Fourier-basis filter parametrization is taken from Xie et al., but the paper applies it in this context and the ablation shows it contributes.\n\nNow the soft spots. The large claim—'rigorous propagation of spatiotemporal rotation symmetry constraints'—does not survive contact with the implementation. The temporal equivariant layer's group SG is stated in the paper to act as a trivial group in practice, with only the identity element. So the temporal layer preserves the spatial group structure; it does not encode a temporal rotation symmetry. Calling this 'temporal rotation equivariance' is a stretch, and the reader's worry lands. Second, Eq. (15) assumes the learned SREC-DC update is rotation-equivariant, but the DC update depends on the sampling operator through A^H(Ax − y). The Cartesian VISTA mask is anisotropic and does not commute with 90° rotations. So even if the network layers are equivariant, the data consistency path is not, and the paper doesn't supply a proof for the learned DC module that ignores this. The stress-test note is right: the global equivariance theorem is unproven. The empirical gains may well come from the extra structure; they just can't be attributed to exact spatiotemporal equivariance as claimed.\n\nThis matters, but it doesn't sink the paper. The experiments are reproducible in the limited sense that the method is clearly specified, though no code or data are released. For a methods paper in an applied journal, I'd want the authors to soften the 'rigorous' language, prove equivariance under the actual acquisition or state the conditions under which it holds, and honestly describe the temporal layer as channel-shift equivariance rather than temporal rotation equivariance. I'd send it to referees; it has enough novelty and enough solid empirical reporting to warrant the time. Just don't let the abstract's second sentence go through as written.","headline":"Strong empirical results on cardiac cine MRI, but the 'rigorous spatiotemporal rotation equivariance' claim outruns the actual construction—worth refereeing, with revision demands.","tokens_in":21852,"tokens_out":3220,"would_cite":true,"duration_ms":34677,"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 deep unrolling network that bakes rotation symmetry into both space and time improves dynamic MRI reconstruction under heavy undersampling.","keywords":["dynamic MRI reconstruction","rotation equivariance","spatiotemporal symmetry","deep unrolling network","cardiac cine MRI","equivariant convolution","data consistency","filter parameterization"],"falsifier":"Train DUN-SRE but replace the temporal equivariant layer's cyclic channel shift with an ordinary 1D convolution of the same receptive field and parameter count; if PSNR, SSIM, and HFEN on the in-house cine dataset at R=20 and R=24 do not degrade, the reported gains do not come from temporal rotation equivariance.","tokens_in":1559,"feed_emoji":"🫀","tokens_out":2988,"duration_ms":87646,"temperature":0.7,"pith_summary":"This paper tries to establish that dynamic MRI reconstruction improves when the network architecture itself enforces rotation symmetry in both space and time, not just within individual frames. It proposes DUN-SRE, a deep unrolling network whose data-consistency and proximal modules are both built from spatiotemporal rotation-equivariant convolutions, so the symmetry constraint propagates through every iteration. On cardiac cine datasets it reports consistent gains over prior methods at all tested acceleration factors, with the largest improvements at high undersampling and in rotation-symmetric structures such as myocardial boundaries and papillary muscles. If the claim holds, this offers a parameter-efficient way to make dynamic reconstruction robust to aggressive undersampling without relying on data augmentation alone.","feed_headline":"Symmetry in space and time lifts cardiac MRI quality","feed_subtitle":"A (2+1)D rotation-equivariant unrolling network beats prior methods at every acceleration factor, R=8-24.","key_machinery":"The load-bearing object is the SREC (Spatiotemporal Rotation Equivariant Convolution) module, a (2+1)D equivariant convolution that separates spatial and temporal processing. Spatial layers apply group-equivariant convolutions over the rotation group G (set to four pi/2 rotations), producing feature maps indexed by rotation; the temporal layer is a pseudo-1D convolution with 1x1 spatial support acting on a group SG that is effectively trivial but still performs cyclic shifts of the group sub-channels, which transfers equivariance between consecutive 2D layers. The same SREC block builds both the proximal mapping network and a learned data-consistency network, so rotation equivariance is preserved through every unrolled iteration. A Fourier-series filter parameterization replaces naive filter rotation to avoid interpolation artifacts, and the paper argues this combined design is what prior unrolling networks lacked.","core_discovery":"The central claim is that rotation equivariance must be modeled jointly in the spatial and temporal dimensions for dynamic MRI, and that this can be done inside an unrolled optimization network. Prior equivariant CNNs cover planar or volumetric rotations, but a naive temporal convolution breaks the group structure established by the 2D layers. DUN-SRE introduces a (2+1)D spatiotemporal rotation equivariant convolution (SREC) that factorizes the operation into 2D group-equivariant spatial layers and a dedicated 1D temporal equivariant layer, with cyclic group-channel shifts maintaining equivariance across time. Because both the proximal network and the data-consistency network use SREC, the entire unrolled network is equivariant end to end, and a Fourier-basis filter parameterization keeps the rotated kernel representations accurate. The paper reports quantitative gains over five prior methods on its in-house cine dataset across R=8, 12, 16, 20, and 24, and visual improvements in the temporal (y-t) domain as well as on the OCMR generalization and prospective data.","pith_inferences":["The practical content of the temporal equivariance is a cyclic shift of rotation-labeled channels across frames, so the physical interpretation as 'rotation over time' is looser than the name suggests; the benefit may come from enforcing group coherence and weight sharing rather than from a true temporal rotation model.","The same (2+1)D equivariant construction should transfer to other dynamic inverse problems, such as video super-resolution or motion-resolved imaging, whenever frames contain orientation-varied repeating structures.","Because the equivariance is discrete (pi/2 rotations), part of the gain may come from matching the four-fold symmetry of Cartesian sampling and the roughly circular heart cross-section; testing N_G=8 or continuous steerable filters would reveal how much finer rotational symmetry adds."],"forward_implications":["Rotation symmetry in dynamic reconstruction is best enforced by architecture rather than by data augmentation, since the unrolled network remains equivariant at every iteration.","The equivariant data-consistency module is load-bearing: ablations show that replacing it with a conventional CNN-based DC term degrades performance and breaks end-to-end equivariance.","Weight sharing across rotation group and time means the symmetry prior improves parameter efficiency, which should help in low-data regimes.","The method generalizes across scanners and to prospectively undersampled data without fine-tuning, suggesting the symmetry prior transfers across acquisition domains."],"supporting_citations":[{"why":"Supplies the Fourier-series basis-function filter parameterization the paper adapts to reduce aliasing in rotated 2D convolutions.","marker":"[30]"},{"why":"Defines group equivariant convolutions and the equivariance of nonlinearities, which the SREC layers build upon.","marker":"[25]"},{"why":"Establishes the design principle that physically meaningful transformation symmetries should be embedded as inductive biases in unrolling networks.","marker":"[53]"},{"why":"Provides a prior rotation-equivariant proximal operator for static image restoration, which the paper extends to both spatial and temporal dimensions.","marker":"[55]"},{"why":"The authors' own preliminary spatiotemporal equivariant CNN used as a proximal network, which DUN-SRE generalizes into a full unrolling framework.","marker":"[35]"},{"why":"Introduces learned data consistency for dynamic MR imaging, which DUN-SRE replaces with an equivariant SREC-based DC network.","marker":"[52]"}],"fun_headline_variants":["Spatiotemporal rotation equivariance sharpens dynamic MRI","Equivariant unrolling network enhances cardiac cine MRI","Joint space-time symmetry improves MRI reconstruction","Rotation symmetry across time lifts dynamic MRI quality","DUN-SRE: symmetry-aware unrolling for dynamic MRI"],"cache_read_input_tokens":23936,"weakest_assumption_plain":"The construction assumes that cardiac dynamics can be faithfully modeled by a rotation-equivariant structure in which the temporal layer only cyclically shifts rotation-labeled channels, rather than requiring a more general motion model.","fun_headline_variants_meta":{"raw":{"variants":["Spatiotemporal rotation equivariance sharpens dynamic MRI","Equivariant unrolling network enhances cardiac cine MRI","Joint space-time symmetry improves MRI reconstruction","Rotation symmetry across time lifts dynamic MRI quality","DUN-SRE: symmetry-aware unrolling for dynamic MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2344,"prompt_tokens":1009,"completion_tokens":1335,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":1275}},"tokens_in":625,"tokens_out":1335,"duration_ms":11495,"temperature":1.0,"reasoning_tokens":1275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:30:16.722135+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DUN-SRE but replace the temporal equivariant layer's cyclic channel shift with an ordinary 1D convolution of the same receptive field and parameter count; if PSNR, SSIM, and HFEN on the in-house cine dataset at R=20 and R=24 do not degrade, the reported gains do not come from temporal rotation equivariance.","supporting_citations":[{"cited_title":"Fourier series expansion based filter parametrization for equivariant convolutions,","cited_arxiv_id":null,"evidence_quote":"Supplies the Fourier-series basis-function filter parameterization the paper adapts to reduce aliasing in rotated 2D convolutions."},{"cited_title":"Equivariant neural networks for inverse problems,","cited_arxiv_id":null,"evidence_quote":"Establishes the design principle that physically meaningful transformation symmetries should be embedded as inductive biases in unrolling networks."},{"cited_title":"Rotation equivariant proximal operator for deep unfolding methods in image restoration,","cited_arxiv_id":null,"evidence_quote":"Provides a prior rotation-equivariant proximal operator for static image restoration, which the paper extends to both spatial and temporal dimensions."},{"cited_title":"SRE- CNN: A spatiotemporal rotation-equivariant CNN for cardiac cine MR imaging,","cited_arxiv_id":null,"evidence_quote":"The authors' own preliminary spatiotemporal equivariant CNN used as a proximal network, which DUN-SRE generalizes into a full unrolling framework."},{"cited_title":"Learning data consistency and its application to dynamic MR imaging,","cited_arxiv_id":null,"evidence_quote":"Introduces learned data consistency for dynamic MR imaging, which DUN-SRE replaces with an equivariant SREC-based DC network."}],"review_version":1}