{"id":"6c47bf62-dbdc-4683-8f90-230c0aa4709c","arxiv_id":"2501.14592","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An equivariant U-Net using symmetric rotation-equivariant kernels improves retinal vessel segmentation on rotated images while using far fewer parameters than baseline networks.","lead":"A U-Net that replaces ordinary convolution kernels with symmetric rotation-equivariant kernels segments retinal blood vessels more accurately and uses 25% of the parameters of a standard U-Net. The method stays stable when images are rotated, which matters for medical images where organs and vessels have no fixed orientation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Performance claim rests on test-set-tuned hyperparameters and noise-level metric differences; significance not established.","rationale":"The reader's weakest assumption focused on the mismatch between exact discrete equivariance and continuous-rotation claims, while also noting the kernel-size selection on the test set and missing error bars. I agree that the small-angle equivariance gap is a real conceptual weakness: the SRE kernel with concentric square bands is exactly equivariant only for the dihedral group D4 (90° rotations and reflections), and the paper provides no analytical or numerical argument that the same kernel remains effectively equivariant for 1°–5° rotations. However, this concern primarily attacks the mechanism, not the raw performance numbers. The more load-bearing issue is the reliability of the empirical comparison itself. The paper selects its proposed kernel configuration by maximizing Dice on the test set (Table 2), and it reports no uncertainty quantification. The reported differences between SRE U-Net and the baselines are tiny relative to the expected sampling noise of a 20-image test set, so 'significantly surpasses' is unsupported. If a proper validation-based selection and significance testing were performed, the empirical claim might still hold, but the current evidence does not establish it. The paper has positive features: the code is released, the architecture change is well specified, and the comparison to a same-architecture U-Net partially isolates the SRE kernel's effect. These strengths make the work promising but the empirical foundation needs additional verification before the central claim can be accepted. Therefore I recommend keeping the reader's CONDITIONAL verdict: the paper should be accepted only if the authors provide statistical validation and use a proper validation set for hyperparameter selection.","tokens_in":6894,"tokens_out":8242,"duration_ms":72614,"concrete_test":"Using the released code, retrain SRE U-Net, U-Net, and FR U-Net with the same recipe and compute per-image Dice scores on the 20 test images. Then calculate paired bootstrap 95% confidence intervals for the mean Dice difference (SRE U-Net minus FR U-Net, and SRE U-Net minus U-Net) at 0° and at ±5°. If either interval includes 0, the 'surpasses' claim is not supported. Additionally, reselect the k configuration on a validation split (e.g., 6 of the 20 training images) rather than the test set, and evaluate the chosen model once on the held-out test set; if the resulting Dice drops below the reported 0.8087 or below FR U-Net's 0.8079, the reported advantage is an artifact of test-set tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that SRE U-Net surpasses U-Net and FR U-Net on DRIVE and does so with 1.7% of the parameters—is not supported by the reported evidence, for two compounding reasons. First, the kernel configuration k=[9,7,5] was selected by maximizing Dice on the same test set used for the final evaluation (Table 2: Dice 0.8064 at ±5° vs 0.7960 for [7,7,7]; the chosen configuration is labelled 'Proposed'). This is test-set-based model selection, which optimistically biases the reported numbers. Second, no error bars, repeated runs, or statistical significance tests are provided. With only 20 test images, a typical per-image Dice SD of 0.05–0.1 gives a standard error of the mean around 0.01–0.02. The differences that carry the headline claim are far smaller: at 0°, SRE U-Net Dice is 0.8087 vs 0.8079 for FR U-Net and 0.7998 for U-Net; at ±5°, 0.8046 vs 0.8027 and 0.7947, respectively. These gaps are within the noise floor, so 'significantly surpasses' and 'best across most metrics' are not established. This is load-bearing because the entire conclusion rests on these empirical comparisons; if the differences vanish under proper evaluation, the parameter-efficiency and rotation-robustness claims lose their evidence base.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SRE U-Net, a U-Net variant in which all convolution kernels are replaced by symmetric rotation-equivariant (SRE) kernels with a small number of trainable parameters, two downsampling stages, and linear-interpolation upsampling. The authors evaluate the method on the DRIVE retinal vessel segmentation dataset, using both original test images and images rotated in 1-degree increments within +/-5 degrees, and compare against standard U-Net, Group U-Net, ES U-Net, FR U-Net, and the Frangi filter. They report that SRE U-Net achieves the best accuracy, specificity, IoU, Dice, and AUC while using only 0.12M parameters, and they claim that it surpasses the state-of-the-art FR U-Net while using 1.7% of its trainable parameters. The paper also includes an ablation over kernel sizes and qualitative difference maps.","tokens_in":7193,"tokens_out":5671,"duration_ms":62682,"significance":"If the empirical claims were established, the paper would provide a useful plug-in convolution kernel that improves rotation robustness and parameter efficiency for segmentation: reducing from 0.48M parameters for the U-Net baseline to 0.12M and from 6.97M for FR U-Net to 0.12M is striking, and the code is made publicly available. The use of a public benchmark, multiple metrics, and a rotation-based test protocol are strengths. However, the central quantitative claims are not currently supported: the kernel configuration is selected on the same test set used for the final evaluation, no error bars or significance tests are reported, and the headline differences are smaller than the likely noise floor. The 90-degree equivariance is also a design consequence of the kernel construction rather than an independently discovered property. The direction is plausible, but the reported margins need substantially stronger evidence.","major_comments":[{"comment":"The kernel configuration k=[9,7,5] is selected by maximizing Dice on the same rotated test set used for the final evaluation: Table 2 reports Dice averaged over +/-5 degrees of 0.8064 for [9,7,5] versus 0.7960 for [7,7,7], and the selected configuration is labelled 'Proposed'. Because the test set contains only 20 images and no error bars are given, this test-set-based model selection makes the final headline numbers optimistically biased. The configuration choice should be made on a held-out validation set, or the authors should report nested cross-validation; without this, the reported margins cannot be taken at face value.","section":"Sec. 3.4, Table 2"},{"comment":"No error bars, repeated runs, or statistical significance tests are reported. The decisive comparisons are very small: at 0 degrees, SRE U-Net Dice is 0.8087 versus 0.8079 for FR U-Net and 0.7998 for U-Net; at +/-5 degrees, the values are 0.8046 versus 0.8027 and 0.7947, respectively. With n=20 test images and per-image Dice variability typical of this dataset, the standard error of the mean is on the order of 0.01-0.02, so the claimed 'significantly surpasses' and 'best across most metrics' are not established. The authors should provide per-image metric distributions, multiple training seeds, or statistical significance tests for the differences.","section":"Sec. 3.2, Table 1"},{"comment":"The strict rotation-equivariance claim is broader than the construction supports. The SRE kernel is invariant under grid-preserving rotations and reflections, such as 90-degree rotations, by construction of the bands, but the paper's central evidence includes continuous small-angle rotations (1 to 5 degrees) where exact equivariance does not follow from Eq. (1). No analytical statement is given for small angles, and the experiments do not isolate the kernel's contribution from the architectural changes (two downsampling stages, linear-interpolation upsampling) or from the kernel-size selection. Please either narrow the equivariance claim to the discrete symmetry group or add an analytical or carefully controlled empirical argument for small-angle robustness.","section":"Sec. 2.1, Eq. (1); Sec. 3.2"}],"minor_comments":[{"comment":"The title contains a typo: 'EQUIV ARIANT' should be 'EQUIVARIANT'.","section":"Title, Sec. 2"},{"comment":"Please clarify whether the standard U-Net baseline and the other equivariant baselines also use two downsampling stages and linear-interpolation upsampling; the text says 'same architecture' but the original U-Net uses transposed convolutions, so the comparison is otherwise not architecture-matched.","section":"Sec. 2.2"},{"comment":"The reported MSE values need a precise definition (mean over pixels, normalized, or image-level) and the qualitative result is shown for one image; a quantitative aggregate over the full test set would be more informative.","section":"Fig. 3"},{"comment":"The band-index construction would be clearer if the dimensions of the binary index matrix I and the reshaping operator psi were defined explicitly in the text; Fig. 1 partially provides this, but the notation in Eq. (1) is otherwise ambiguous.","section":"Sec. 2.1, Eq. (1)"},{"comment":"The claim that SRE-Conv can be 'easily integrated into other existing segmentation frameworks' is speculative and should be framed as future work rather than an established result.","section":"Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is essentially an application of the SRE-Conv kernel from the same group's prior preprint [18] to a U-Net for vessel segmentation. The novelty lies in the architectural integration and the DRIVE evaluation, so the empirical claims carry most of the weight. Given the test-set-based kernel selection and the absence of significance testing, the current results cannot support the paper's strong 'surpasses' statements. The authors should be asked to re-run with a proper validation-based configuration selection and statistical comparison; if those results still show a consistent advantage, the paper could be a solid engineering contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: SRE U-Net is a clean, plausible application of the authors' own SRE-Conv kernel to segmentation, and the parameter-efficiency story is real—0.12M parameters versus 6.97M for FR U-Net is a big difference, not a rounding error. What is not established is the headline performance claim. The Dice gaps driving \"surpasses\" are 0.001 to 0.01 on a 20-image test set, with no error bars and no significance testing. The kernel-size configuration was chosen by looking at Dice on the same test set used for the final numbers. That is test-set tuning, and the gaps are within the noise floor. I think the reader's strongest concern lands.\n\nThe paper does several things well. It makes a sensible architectural choice: two downsamplings reduce parameters, and linear-interpolation upsampling instead of trainable up-convolution is the right way to preserve equivariance through the decoder. The benchmarking includes reasonable baselines, a range of metrics, and a qualitative consistency check under 90-degree and 45-degree rotation. The code link is a positive, though I did not run it and cannot verify the reported numbers from the repo alone.\n\nThe soft spots are real and load-bearing for the headline claims. First, no repeated runs, error bars, or statistical tests appear anywhere. With 20 test images, a per-image Dice SD of 0.05–0.1 puts the standard error around 0.01–0.02; the SRE U-Net versus FR U-Net differences at 0 and ±5 degrees are smaller than that. So \"significantly surpasses\" is not supported. Second, the ablation in Table 2 selects k=[9,7,5] by maximizing Dice on the test set, and that same configuration then produces the paper's main results. That is optimistic selection, not a fair evaluation. Third, the small-angle robustness story is underspecified: the kernel's exact equivariance is for grid-preserving rotations like 90 degrees and reflections, while the continuous 1–5 degree results are empirical and confounded by nearest-neighbor resampling. The paper never isolates the SRE kernel's contribution from the other architectural changes, so we do not know how much of the gain is the kernel rather than the smaller network or the interpolation upsampling.\n\nNone of this makes the central idea wrong. A parameter-efficient radially symmetric convolution kernel is a sensible plug-in for medical segmentation, and the qualitative rotation consistency is visible. But the empirical case needs a validation split for the ablation, error bars with significance tests on the rotated test set, and an explicit statement about what exact equivariance the kernel does and does not provide.\n\nWho is this for? People working on rotation-robust medical image segmentation and small-model equivariant CNNs. I would not desk-reject it, but I would send it to peer review with a strong request for proper statistical evaluation. As it stands, the contribution is promising but the reported margins are not yet credible.","headline":"A plausible plug-in for rotation-robust U-Net segmentation, but the headline performance margins are within the noise floor and need a proper evaluation before they carry weight.","tokens_in":7668,"tokens_out":2807,"would_cite":false,"duration_ms":29346,"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":"This paper claims that replacing standard convolutions with symmetric rotation-equivariant kernels makes a U-Net segment retinal vessels accurately on rotated inputs while using 0.12M parameters, about 1.7% of a 6.97M-parameter…","keywords":["retinal vessel segmentation","rotation equivariance","symmetric convolution kernel","U-Net","parameter efficiency","fundus imaging","semantic segmentation","convolutional neural networks"],"falsifier":"Train a comparison U-Net with the identical architecture and kernel-size schedule as SRE U-Net but standard asymmetric convolutions, matched for parameter count, and evaluate both on the DRIVE test set rotated in 1-degree steps. If the standard network matches SRE U-Net's Dice and AUC under rotation, the symmetric kernel is not what carries the reported robustness.","tokens_in":6724,"feed_emoji":"👁","tokens_out":6215,"duration_ms":49743,"temperature":0.7,"pith_summary":"The paper argues that a U-Net whose ordinary convolution kernels are replaced by symmetric rotation-equivariant kernels can segment retinal vessels accurately even when test images are rotated, while using far fewer trainable parameters than standard or state-of-the-art networks. On the DRIVE dataset, the proposed SRE U-Net reports the best accuracy, specificity, IoU, Dice, and AUC among the compared methods, and it outperforms a 6.97M-parameter state-of-the-art segmentation network using only 1.7% of that parameter count. The core idea is that rotational equivariance can be built directly into a standard segmentation architecture through a cheap kernel constraint, rather than through large group-convolution or steerable-filter stacks. If the paper is right, rotation-robust segmentation becomes practical in data- and compute-limited clinical settings, where image orientation is often inconsistent or unknown.","feed_headline":"Symmetric kernels make U-Net rotation-stable at 0.12M parameters","feed_subtitle":"On retina vessel data, a 0.12M-parameter model matches a 6.97M-parameter state-of-the-art under small rotations.","key_machinery":"The load-bearing object is the symmetric rotation-equivariant convolution kernel (SRE-Conv kernel). It partitions a $k \\times k$ kernel into $b = \\lfloor k/2 \\rfloor + 2$ concentric bands; a binary index matrix $I$ assigns each kernel entry to one band, and the kernel is reconstructed as $\\psi(\\Theta_i \\cdot I)$, where $\\Theta_i$ holds only $b$ trainable values per channel. This construction forces the kernel to be centrally symmetric, which makes the convolution commute with 90-degree rotations and reflections and, the paper argues, generalizes to finer rotations through the learned band structure. Two architectural choices complete the mechanism: every convolution in the encoder and decoder is replaced with SRE-Conv, and decoder upsampling uses linear interpolation instead of trainable strided up-convolution, because strided up-convolution would break equivariance under rotation.","core_discovery":"On the paper's own terms, the central discovery is that applying SRE-Conv, a convolution kernel constrained to be centrally symmetric and encoded by a small number of band parameters, throughout a U-Net yields a segmentation model that is rotation- and reflection-equivariant in its feature maps and stays accurate when inputs are rotated by small angles (1 to 5 degrees) as well as by 90 degrees. In the reported experiments, the SRE U-Net (0.12M parameters) attains the highest averaged accuracy (0.9600), specificity (0.9870), IoU (0.6795), Dice (0.8087), and AUC (0.9767) among U-Net, Group U-Net, ES U-Net, FR U-Net, and Frangi baselines, with FR U-Net slightly ahead on sensitivity. The authors take this as evidence that symmetric kernels give the network both parameter efficiency and rotational robustness, and that the SRE kernel can serve as a drop-in replacement for the convolution layers in segmentation backbones.","pith_inferences":["The exact equivariance guarantee holds for grid-preserving rotations such as 90 degrees and reflections; the observed robustness to continuous 1-to-5-degree rotations is an empirical property that likely depends on the band structure, interpolation, and data augmentation together, not on a proven symmetry for arbitrary angles.","A cleaner test of the kernel's contribution would hold every architectural choice fixed and vary only kernel symmetry; the paper changes downsampling depth and upsampling method at the same time, so part of the gain may come from those changes rather than the symmetric kernel alone.","The same kernel could be evaluated on histopathology or other orientation-agnostic imaging where vessels and tissue structures appear at arbitrary angles; a public multi-dataset benchmark would show whether the parameter-efficiency advantage persists outside DRIVE."],"forward_implications":["Retinal vessel segmentation can become rotation-robust without the memory overhead of group convolution: SRE U-Net uses 0.12M parameters versus 1.92M for Group U-Net and 0.72M for ES U-Net in the comparison.","Because the SRE kernel is designed as a drop-in replacement for standard convolutions, the same equivariance benefit could transfer to other U-Net-based segmentation tasks where object orientation is arbitrary.","A smaller, cheaper segmentation model makes rotation-equivariant learning usable in settings with limited training data or GPU memory.","The reported stability under small rotations implies fewer inconsistent predictions when clinical fundus images are acquired at slightly different angles."],"supporting_citations":[{"why":"Supplies the SRE-Conv kernel implementation and its parameter-efficient band formulation.","marker":"[18]"},{"why":"Provides the U-Net backbone architecture that SRE U-Net modifies.","marker":"[1]"},{"why":"Supplies the DRIVE dataset, the evaluation benchmark for vessel segmentation.","marker":"[19]"},{"why":"State-of-the-art retinal vessel segmentation network used as the performance baseline (6.97M parameters).","marker":"[20]"},{"why":"Group-equivariant U-Net baseline compared for rotation equivariance.","marker":"[10]"},{"why":"Steerable E(2)-equivariant CNN methods that define the ES U-Net baseline.","marker":"[14, 15]"},{"why":"Frangi filter-based non-deep-learning vessel segmentation baseline.","marker":"[21]"}],"fun_headline_variants":["Symmetric U-Net: rotation-equivariant vessels at 0.12M params","Param-efficient U-Net with rotation equivariance for vessels","Vessel segmentation that stands up to rotations with 0.12M params","SRE-Conv U-Net: fewer parameters, steady under rotation","Rotation-stable vessel segmentation with symmetric kernels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the improved accuracy and rotation stability come from the symmetric kernel's equivariance, rather than from the smaller model, the two-downsampling architecture, the interpolation upsampling, or from configuring kernel sizes on the test set.","fun_headline_variants_meta":{"raw":{"variants":["Symmetric U-Net: rotation-equivariant vessels at 0.12M params","Param-efficient U-Net with rotation equivariance for vessels","Vessel segmentation that stands up to rotations with 0.12M params","SRE-Conv U-Net: fewer parameters, steady under rotation","Rotation-stable vessel segmentation with symmetric kernels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1565,"prompt_tokens":985,"completion_tokens":580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":489}},"tokens_in":601,"tokens_out":580,"duration_ms":5614,"temperature":1.0,"reasoning_tokens":489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:59:38.924730+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a comparison U-Net with the identical architecture and kernel-size schedule as SRE U-Net but standard asymmetric convolutions, matched for parameter count, and evaluate both on the DRIVE test set rotated in 1-degree steps. If the standard network matches SRE U-Net's Dice and AUC under rotation, the symmetric kernel is not what carries the reported robustness.","supporting_citations":[{"cited_title":"General e (2)- equivariant steerable cnns,","cited_arxiv_id":null,"evidence_quote":"Supplies the SRE-Conv kernel implementation and its parameter-efficient band formulation."},{"cited_title":"Improved Vessel Segmentation with Symmetric Rotation-Equivariant U-Net","cited_arxiv_id":"2501.14592","evidence_quote":"Provides the U-Net backbone architecture that SRE U-Net modifies."},{"cited_title":"A program to build E(N)-equivariant steerable CNNs,","cited_arxiv_id":null,"evidence_quote":"Supplies the DRIVE dataset, the evaluation benchmark for vessel segmentation."},{"cited_title":"Use of sym- metric kernels for convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"State-of-the-art retinal vessel segmentation network used as the performance baseline (6.97M parameters)."},{"cited_title":"Rotation equivari- ant and invariant neural networks for microscopy image analysis,","cited_arxiv_id":null,"evidence_quote":"Group-equivariant U-Net baseline compared for rotation equivariance."},{"cited_title":"Rotated ring, radial and depth wise separable radial convolutions,","cited_arxiv_id":null,"evidence_quote":"Frangi filter-based non-deep-learning vessel segmentation baseline."}],"review_version":1}