{"id":"5328bdfd-08ef-4421-ace2-1c09749a3a6a","arxiv_id":"2502.02418","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A deep-learning CT reconstruction framework that enforces cyclic rotation equivariance between sinogram and image domains shows improved generalization on brain phantoms, though the underlying equivariance property is classical.","lead":"This paper argues that tomography systems with N-fold rotational symmetry give a cyclic group equivariance between rotated projections and rotated images, and builds this into a deep-learning reconstruction network using a cyclic convolution layer and a symmetry regularizer. Tests on digital brain phantoms show better stability and fewer artifacts than a plain learned reconstruction baseline, especially when training data differ from test data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bicubic rotation surrogate does not form a group action, so CGER and Fig. 9 may measure consistency with a biased operator rather than true C_N equivariance.","rationale":"The reader's weakest assumption points to the bicubic interpolation proxy, and I agree that this is the least secure link. The theory itself, Eqs. (5)-(11), is a valid continuous-domain derivation, and CREC with circular padding is a reasonable discrete implementation of angular equivariance. The problem is the discrete image-domain rotation used both as the CGER target and as the evaluation metric. If the surrogate is biased, Eq. (11)'s practical instantiation fails, and the central empirical claim that enforcing equivariance improves stability and generalization is no longer cleanly supported. Several other limitations (trivial CREC kernel width 1, no parameter-matched baseline, fuzzy test sets from the same brain models) are real but secondary; the interpolation surrogate is more directly load-bearing because it affects the definition of the loss and the metric used to validate the central claim. I therefore keep the reader's conditional verdict: the concern is concrete and testable but not yet demonstrated to invalidate the paper.","tokens_in":13376,"tokens_out":8064,"duration_ms":77962,"concrete_test":"Use FBP as an approximately C_N-equivariant reference (or a U-Net trained only with 90-degree augmentations), and on the discrete test set compute the CGER error 1/127 Σ_i ||F(T_i g) - R_i(F(g))||_2 with bicubic R_i for i=1..127. If a near-equivariant model shows large, angle-dependent error at non-90-degree angles, the surrogate is biased. Also measure group-composition error ||R_i(R_j(x)) - R_{i+j}(x)||_2 for random x; if this error is comparable to the CGER loss reductions claimed in Table I, the reported equivariance improvements are dominated by interpolation artifacts rather than true symmetry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central equivariance statement, Eq. (11), is exact only for continuous rotations R_m of a continuous object, whereas the network outputs a discrete pixel image and CGER (Eq. 14) plus the Fig. 9 evaluation realize R_m by bicubic interpolation on that pixel grid. Bicubic rotation is not a group representation of C_N: it is not exactly invertible, and composition R_i∘R_j only approximates R_{i+j} after resampling. Thus the regularizer and the symmetry-consistency metric measure consistency with an approximate, angle-dependent surrogate rather than with true C_N equivariance. This is load-bearing because the reported CGER gains (e.g., SSIM +0.31 for CSF) and the 'more uniform' polar plots in Fig. 9 could largely reflect fitting this surrogate. The paper offers no calibration of the surrogate: no comparison with exact discrete rotations (e.g., Fourier shift or shear-based), no check that a known equivariant method such as FBP has small CGER error, and no measurement of composition error R_i R_j versus R_{i+j}. The anomalies noted at 90/180/270 in Fig. 9 reinforce the concern, since those are the angles where interpolation should be most reliable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that tomography systems with N-fold rotational symmetry induce a cyclic group C_N that acts by cyclic rotation on projection data and by discrete rotation on the reconstructed image. The central theoretical claim, Eq. (11), states that any reconstruction map F must be equivariant under this group, i.e., F[T_m g] = R_m[y] for all m in Z_N. Based on this, the authors propose two mechanisms to enforce the symmetry in a deep-learning reconstruction network: cyclic rotation equivariant convolution (CREC) layers applied to the projection domain, and a cyclic group equivariance regularization (CGER) term that penalizes deviations from equivariance during training. They integrate these modules into an AUTOMAP-style domain-transform network and evaluate on digital brain phantoms, training on discrete phantoms and testing on fuzzy phantoms from the same brain models. The results show that the modified network with CREC and CGER yields higher SSIM, better artifact suppression, and more uniform angular consistency in polar plots than the unmodified AUTOMAP baseline.","tokens_in":13592,"tokens_out":3528,"duration_ms":35339,"significance":"If the central claim is correct, the paper offers a principled, hardware-derived symmetry constraint that could improve generalization and data efficiency of deep-learning tomographic reconstruction, which is a timely and practically relevant contribution. The derivation from sensitivity functions to Eq. (11) is self-contained and does not rely on fitted parameters, which is a clear strength. The experimental design includes a meaningful distribution shift (discrete to fuzzy phantoms) and reports notable quantitative gains. However, the paper does not release code or data, does not control for parameter count in ablations, and the implementation of the proposed CREC layer appears degenerate when the kernel width is 1. The symmetry-consistency evaluation and the CGER regularizer both depend on a bicubic-interpolation surrogate for rotations that is not an exact group action, so the quantitative gains may partly reflect fitting this surrogate rather than true C_N equivariance. These issues must be addressed before the claims can be fully accepted.","major_comments":[{"comment":"The CGER regularizer and the symmetry-consistency evaluation in Fig. 9 realize the image-domain rotations R_i with bicubic interpolation. Bicubic rotation of a discrete pixel grid is not a group representation of C_N: it is not exactly invertible, and composition R_i after R_j only approximates R_{i+j} due to resampling. Thus the regularizer enforces consistency with an angle-dependent approximate surrogate rather than with true C_N equivariance. The paper acknowledges interpolation anomalies at 90/180/270 degrees but does not calibrate the surrogate error. Please provide a quantitative measure of the surrogate's deviation from a true group action (e.g., composition error ||R_i(R_j(x)) - R_{i+j}(x)|| for representative images), compare with exact discrete rotation implementations such as Fourier-domain rotation or shear-based rotation, and report the CGER error for a known equivariant method like FBP. Without such calibration, the reported SSIM gains and the 'more uniform' polar plots in Fig. 9 may largely reflect fitting the surrogate.","section":"Implementation Details and Eq. (14) (CGER)"},{"comment":"The paper states in the CREC design section that the kernel width is set to 1, eliminating the need for wrapping around. A kernel of width 1 along the angular axis performs no mixing of neighboring angular positions, so the cyclic equivariance is trivially realized but the layer does not exploit the circular structure of the projection domain as claimed. This makes the contribution of the CREC layer ambiguous: the observed improvements of Mod-AUTOMAP over AUTOMAP could stem from the added fully-connected or convolutional parameters rather than from cyclic equivariance. Please clarify the exact kernel size used in the experiments, and if the width is indeed 1, explain how this demonstrates a cyclic equivariant convolution design; otherwise, provide results with a non-trivial kernel width and report the kernel size.","section":"CREC layer, Fig. 2 and Implementation Details"},{"comment":"The comparison between AUTOMAP and Mod-AUTOMAP (with three CREC layers) is not an ablation that controls for parameter count or model capacity. Adding three layers increases the number of trainable parameters and can improve performance independently of equivariance. To support the claim that the gains are due to cyclic equivariance, the authors should include a control model with the same architectural changes but without the cyclic constraint (e.g., replacing CREC layers with standard convolutional layers of identical shape and parameter count, or adding a comparable-capacity autoencoder branch). This is load-bearing because the headline results—e.g., SSIM improvement from 0.8506 to 0.9144 on discrete phantoms—could simply reflect a larger network.","section":"Experimental comparisons in Table I"},{"comment":"The fuzzy phantom test sets are derived from the same two brain models as the discrete test set (as stated in the Datasets section). While the pixel-intensity distributions differ (discrete vs. fuzzy), the underlying anatomy is identical, so the 'external test' evaluates a restricted form of distribution shift, not anatomical variability. This limits the strength of the generalization conclusions. Please either add tests on fuzzy phantoms from held-out brain models not used in training or clearly moderate the claims of generalization to new patient anatomies.","section":"Datasets and generalization claims"}],"minor_comments":[{"comment":"Equation (14) is used twice: once for the cyclic rotation convolution definition and once for the CGER regularized objective. Please renumber the equations to avoid ambiguity.","section":"Equation numbering"},{"comment":"There are several typographical errors, e.g., 'Equitation (10)' should be 'Equation (10)', and '𝒚𝒚( 𝒓𝒓′)' should be '𝒚𝒚(𝐫𝐫′)' in the derivation. Please proofread the manuscript.","section":"Typos"},{"comment":"Reference [23] (Hashemi et al.) is listed with the placeholder arXiv number 2301.00000, which appears to be a dummy identifier. Please update to the actual arXiv reference or remove the placeholder.","section":"References"},{"comment":"The polar plots in Fig. 9 are said to show rRMSE across 127 angular transformations, but the axis labels are not visible in the caption and the plots appear to be on different scales. Please add a shared legend and consistent radial-axis scaling so that 'more uniform' and 'lower' errors can be compared across panels.","section":"Figure 9"},{"comment":"The manuscript does not state whether the code, trained models, or phantom datasets will be made available. Given the reproducibility standards of the field, please include a data/code availability statement.","section":"Code and data availability"}],"recommendation":"major_revision","confidential_remarks":"The core theoretical derivation (Eq. (11)) appears sound and is a useful contribution. However, the experimental section does not yet establish that the observed improvements come from the proposed equivariance mechanisms rather than from added capacity or from fitting an approximate rotation surrogate. The authors should be asked to provide parameter-controlled ablations, clarify the CREC kernel implementation, and quantitatively validate the rotation surrogate. If these points are addressed, the paper could be suitable for publication. I also note that the fuzzy test sets are not independent anatomically, which should be made explicit in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core mathematical claim is correct but classical: Eq. (11) restates the rotational equivariance of the Radon transform. What's actually new is applying cyclic group convolutions and a symmetry regularizer to a domain-transform network, and testing it on a deliberate train/test distribution shift. The derivation from sensitivity functions is self-contained and sound, and the experimental scheme—train on discrete phantoms, test on fuzzy ones—is a meaningful robustness probe. The results are consistent: both CREC and CGER help, and the gains on the shifted set are the most striking. The paper is honest about its simplifications, including the kernel-width-1 CREC implementation and the interpolation anomalies at 90/180/270 in Fig. 9.\n\nThe soft spots are real. No code or data is released, which is a serious limitation for an empirical paper. The CREC layer with kernel width 1 along the angle axis is nearly trivial—it does not mix neighboring angles, so 'cyclic equivariant convolution' becomes a per-angle reweighting, and the comparison to AUTOMAP does not control parameter count. The bigger issue is the rotation surrogate: the CGER regularizer and the Fig. 9 symmetry metric both use bicubic interpolation to represent R_m, and bicubic rotation is not a group representation—it resamples, it is not exactly invertible, and composition only approximates the group law. So the symmetry-consistency evaluation partly measures consistency with this interpolation surrogate rather than true C_N equivariance. The reconstruction quality numbers (SSIM/PSNR against ground truth) are not contaminated by this, so the practical gains are credible, but the equivariance interpretation is weaker than claimed. The paper would be stronger with a calibration check: compare against exact discrete rotations (e.g., Fourier shift), or show that a known equivariant method like FBP has small CGER error.\n\nThis deserves a serious referee: the idea is plausible, the derivation is correct, and the experiments are honest. The authors should be asked for code, parameter-matched baselines, and a sanity check of the rotation surrogate before the claims are established. For me, it's a useful read for anyone working on equivariance in image reconstruction, and the distribution-shift test design is worth borrowing.","headline":"A plausible equivariance-based improvement for domain-transform reconstruction, with a sound derivation but experimental gaps and an interpolation caveat that keep me from fully trusting the symmetry claims.","tokens_in":14111,"tokens_out":2891,"would_cite":false,"duration_ms":26133,"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 derives a cyclic equivariance identity coupling rotation of projections to rotation of images, and shows that enforcing it in a deep network improves tomographic reconstruction under distribution shift.","keywords":["cyclic group equivariance","image reconstruction","computed tomography","deep learning","rotational symmetry","equivariant convolution","domain transform","generalization"],"falsifier":"Retrain the CGER-constrained model with an exact discrete rotation (e.g., Fourier-based or three-shear rotation) in place of bicubic interpolation and replot the polar symmetry-consistency error; if the error spikes at 90, 180, and 270 degrees persist, the residual inconsistency is not interpolation noise and the discrete-grid equivariance identity itself is the weak point.","tokens_in":13163,"feed_emoji":"🩻","tokens_out":10059,"duration_ms":88870,"temperature":0.7,"pith_summary":"This paper tries to establish that the rotational symmetry of tomography hardware should be treated as a hard constraint on deep reconstruction networks, not just a property of the data. It derives the equivariance condition that cyclically shifting the angular index of the projection data must rotate the reconstructed image by the corresponding discrete angle, and it identifies the cyclic group underlying both domains. To enforce the condition, the paper contributes two mechanisms: cyclic rotation equivariant convolutions in the projection domain and a cyclic group equivariance regularizer over the full network. On digital brain phantoms, these additions stabilize training and improve reconstruction quality specifically when the test data look different from the training data. If the claim holds, cyclic group equivariance becomes a unifying design principle for any rotationally symmetric imaging system.","feed_headline":"Cyclic rotation equivariance stabilizes deep tomography reconstruction","feed_subtitle":"Respecting the scanner's N-fold rotation cuts artifacts and improves generalization under distribution shift.","key_machinery":"The load-bearing object is the cyclic group $C_N$ together with the equivariance identity that couples its two representations: cyclic shifts $\\mathcal{T}_m$ on the sinogram's angular axis and discrete rotations $R_m$ on the image plane. The CREC layer realizes group convolution on the angular axis, wrapping kernels around the boundary so that the projection feature maps remain circular. The CGER regularizer adds a term summing $\\lVert \\mathcal{F}(\\mathcal{T}_i \\mathbf{g}, \\theta) - R_i \\mathcal{F}(\\mathbf{g}, \\theta) \\rVert^2$ for $i$ in a subgroup $C_M$, balancing it dynamically against the reconstruction loss. Together they cover the two ways a symmetry can enter a network: by architectural design in the projection domain and by optimization across the whole reconstruction map.","core_discovery":"The central claim is equation (11): for every $m \\in \\mathbb{Z}_N$, $\\mathcal{F}[\\mathcal{T}_m \\mathbf{g}] = R_m[\\mathbf{y}]$, where $\\mathcal{F}$ is the reconstruction map, $\\mathcal{T}_m$ cyclically rotates the angular axis of the projection data, and $R_m$ rotates the reconstructed discrete image by $2\\pi m/N$. The paper argues this identity follows from the forward model when the $N$ detector arrays are identical and uniformly distributed around the central axis, and that it equips the projection space with a uniformly distributed circular structure. It then shows how to enforce the identity in practice: CREC layers perform group convolution along the angular axis with circular boundary handling so that projection features inherit the symmetry, and CGER penalizes deviations from the equivariance relation, computed over a subgroup $C_{32}$ of $C_{128}$ with image rotations performed by bicubic interpolation. Across discrete and fuzzy phantom experiments, the constrained networks exhibit lower rotational inconsistency and better out-of-distribution reconstruction than the unconstrained baseline.","pith_inferences":["Beyond the paper: because the equivariance identity is derived purely from detector geometry, the same CREC and CGER modules should transfer to PET, SPECT, and optoacoustic tomography, and the natural test is whether they reduce the number of training angles needed for a target image quality.","Beyond the paper: the regularizer is evaluated only on a subgroup $C_{32}$ of $C_{128}$; sweeping over all divisors of $N$ would reveal how much of the group's symmetry is responsible for the stability gain and whether full-group regularization is worth its extra cost.","Beyond the paper: the observed interpolation anomalies at 90, 180, and 270 degrees suggest replacing the bicubic rotation with an exact discrete rotation operator; this is a directly testable modification that could make the regularizer an exact symmetry constraint rather than an approximate one."],"forward_implications":["A domain-transform network with CREC layers outperforms its unconstrained counterpart on in-distribution discrete phantoms, improving SSIM from 0.8506 to 0.9144 while keeping rRMSE below filtered back-projection.","Adding CGER sharply improves out-of-distribution reconstruction, with SSIM rising by up to 0.3114 for the unmodified network and by up to 0.1862 for the CREC-modified network on fuzzy phantoms.","CGER narrows the training-validation gap and makes convergence smoother, which is a sign of reduced overfitting in the limited-data setting.","Under a full $C_{128}$ symmetry-consistency check, both mechanisms lower the rotation-error distribution, and the CREC-plus-CGER model gives the most uniform errors across all 127 non-identity rotations."],"supporting_citations":[{"why":"Supplies the domain-transform reconstruction baseline that the paper modifies with CREC and CGER.","marker":"[3]"},{"why":"Defines group convolution, the mathematical construction the CREC layer adapts to the cyclic group.","marker":"[14]"},{"why":"Gives the theoretical generalisation benefit of equivariant models, motivating why enforcing equivariance improves out-of-distribution performance.","marker":"[17]"},{"why":"Provides the digital brain phantoms, including the fuzzy variants used to measure distribution-shift performance.","marker":"[27]"},{"why":"Documents stability failures in deep reconstruction, the problem the regularizer targets.","marker":"[12]"},{"why":"Represents existing equivariant-inverse-problem methods that the paper extends from image-domain symmetry to the projection domain.","marker":"[25]"}],"fun_headline_variants":["Cyclic group symmetry cuts artifacts in rotationally symmetric tomography","Equivariant deep learning stabilizes CT reconstruction under distribution shift","Cyclic-equivariant convolution improves tomographic reconstruction with limited data","Scanner symmetry as a prior for robust deep CT reconstruction","Enforcing the N-fold symmetry of CT scanners in deep networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that rotating a reconstructed image with bicubic interpolation is a faithful discrete stand-in for the continuous rotation $R_m$; for angles far from multiples of 90 degrees, interpolation error could be misread as an equivariance violation and distort the reconstruction.","fun_headline_variants_meta":{"raw":{"variants":["Cyclic group symmetry cuts artifacts in rotationally symmetric tomography","Equivariant deep learning stabilizes CT reconstruction under distribution shift","Cyclic-equivariant convolution improves tomographic reconstruction with limited data","Scanner symmetry as a prior for robust deep CT reconstruction","Enforcing the N-fold symmetry of CT scanners in deep networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001321,"raw_usage":{"total_tokens":5401,"prompt_tokens":987,"completion_tokens":4414,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":4331}},"tokens_in":603,"tokens_out":4414,"duration_ms":29515,"temperature":1.0,"reasoning_tokens":4331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:12:44.722034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the CGER-constrained model with an exact discrete rotation (e.g., Fourier-based or three-shear rotation) in place of bicubic interpolation and replot the polar symmetry-consistency error; if the error spikes at 90, 180, and 270 degrees persist, the residual inconsistency is not interpolation noise and the discrete-grid equivariance identity itself is the weak point.","supporting_citations":[{"cited_title":"Image reconstruction by domain- transform manifold learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the domain-transform reconstruction baseline that the paper modifies with CREC and CGER."},{"cited_title":"Geometric deep learning and equivariant neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines group convolution, the mathematical construction the CREC layer adapts to the cyclic group."},{"cited_title":"Provably strict generalisation benefit for equivariant models,","cited_arxiv_id":null,"evidence_quote":"Gives the theoretical generalisation benefit of equivariant models, motivating why enforcing equivariance improves out-of-distribution performance."},{"cited_title":"Twenty new digital brain phantoms for creation of validation image data bases,","cited_arxiv_id":null,"evidence_quote":"Provides the digital brain phantoms, including the fuzzy variants used to measure distribution-shift performance."},{"cited_title":"On instabilities of deep learning in image reconstruction and the potential costs of AI,","cited_arxiv_id":null,"evidence_quote":"Documents stability failures in deep reconstruction, the problem the regularizer targets."},{"cited_title":"Equivariant neural networks for inverse problems,","cited_arxiv_id":null,"evidence_quote":"Represents existing equivariant-inverse-problem methods that the paper extends from image-domain symmetry to the projection domain."}],"review_version":1}