{"id":"2d59462c-24d1-49ec-941b-967b0ecc395f","arxiv_id":"2501.02788","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Adding roughly 15 to 30 learnable Gabor and LoG filter parameters to a Swin-Unet patch embedding improves Dice scores by about one point on two medical segmentation benchmarks.","lead":"GLoG-CSUnet adds a few learnable Gabor and Laplacian-of-Gaussian filters to a Swin-Transformer segmentation network, aiming to capture texture and edges that transformers miss. The authors report Dice score improvements of about one point over existing models on two medical image datasets, at a claimed cost of only 15 to 30 extra parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1.14% and 0.91% Dice gains are within plausible run-to-run variability: no seeds or error bars are reported, baselines are quoted from prior papers, and ablation gaps are the same size as the headline improvement.","rationale":"The stress-test pass identifies the same load-bearing concern as the reader's weakest_assumption: the reported Dice improvements are not distinguished from run-to-run variability because no seeds or error bars are given and baselines are quoted from prior publications with potentially different training protocols. This concern directly challenges the abstract's 'significant improvements' claim, which is the central contribution. The ablation study reinforces the concern, since the differences between the ablations are comparable in size to the headline gain and are each presumably from a single run. I agree with the reader that conditional acceptance is appropriate: the architectural idea is reasonable and cheap to test, but the evidence as presented is insufficient to establish the claimed effect. The parameter-overhead inconsistency is noted but is secondary to the performance-evidence problem; it would only further weaken the 'minimal overhead' selling point if confirmed. Therefore the reader's verdict of CONDITIONAL remains appropriate, and no verdict adjustment is needed.","tokens_in":7402,"tokens_out":3946,"duration_ms":40009,"concrete_test":"Obtain the released code (or reimplement if unavailable) and train CS-Unet and GLoG-CSUnet on Synapse and ACDC under identical conditions: 300 epochs, batch size 24, AdamW, same learning-rate schedule, and at least 5 random seeds per method. Report mean and standard deviation of DSC for each method and compute a paired test per dataset. If the mean gap is within one pooled standard deviation or the 95% confidence interval includes zero, the claimed significant improvement is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GLoG-CSUnet beats CS-Unet by 1.14 percentage points on Synapse and 0.91 on ACDC. All baseline numbers in Tables I and II are copied from earlier publications, while the proposed model is trained from scratch for 300 epochs with dataset-specific learning rates (Section II.E). No method is evaluated with multiple random seeds or error bars, so the reported gap cannot be separated from seed-to-seed variation or from protocol differences such as augmentation, learning-rate schedule, and whether pretrained weights were used (the paper states baselines used pretrained weights if provided). The ACDC ablation (Section III.A) reports Gabor-only 92.01%, LoG-only 91.83%, and combined 92.28%; these component differences of 0.18 and 0.45 percentage points are the same order of magnitude as the claimed headline gain, so a single run per variant cannot establish that the effect is systematic. A secondary inconsistency is the parameter-overhead claim of only 15 and 30 additional parameters, which is difficult to reconcile with Figure 1(b), where filter outputs are concatenated and fed into convolutional layers, increasing the first conv layer's input channels and hence adding thousands of parameters; this is worth checking, but the performance claim is the more load-bearing issue because the paper's contribution rests on the asserted improvement being real.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GLoG-CSUnet, an extension of the Convolutional Swin-Unet (CS-Unet) architecture that inserts learnable Gabor and Laplacian-of-Gaussian (LoG) filters into the patch-embedding stage of a Swin-Transformer-based segmentation model. The method is evaluated on the Synapse multi-organ CT and ACDC cardiac MRI benchmarks. The authors report Dice similarity coefficients of 83.36% on Synapse and 92.28% on ACDC, outperforming CS-Unet by about 1 percentage point, with a claimed overhead of only 15 and 30 additional parameters. An ablation on ACDC reports Gabor-only 92.01%, LoG-only 91.83%, and combined 92.28%. The paper argues that radiomics-inspired filters complement the global modeling of Transformers, particularly for small medical datasets.","tokens_in":7690,"tokens_out":3662,"duration_ms":36652,"significance":"If the reported gains are reproducible, the contribution is a simple and modular patch-embedding enhancement that yields consistent, albeit small, Dice improvements on two standard benchmarks at negligible parameter cost. The paper provides an ablation study and a public code repository, which are useful for verification. However, the evidence as presented is not yet sufficient to establish the central claim: all baseline numbers are quoted from prior publications, no error bars or multiple-seed results are reported, and the headline numbers in the abstract and body are internally inconsistent. The parameter-overhead claim also appears incompatible with the architecture description. The idea is worth publishing if these load-bearing issues are resolved with additional experiments and corrected reporting.","major_comments":[{"comment":"The claimed Dice gains of about 1.14 and 0.91 percentage points over CS-Unet are not supported as 'significant improvements' because no method is evaluated with multiple random seeds or error bars. All baseline numbers in Tables I and II are copied from earlier publications, whereas GLoG-CSUnet is trained from scratch for 300 epochs with dataset-specific learning rates, augmentation, and a batch size of 24; differences in training protocol can easily produce differences of this magnitude. The ablation in Section III.A compounds the problem: the component gaps (Gabor-only 92.01 vs LoG-only 91.83 vs combined 92.28) are the same order of magnitude as the headline gain, so a single run per variant cannot establish that the effect is systematic. Please provide mean and standard deviation over at least three seeds for the proposed method, and ideally re-run CS-Unet under the identical training protocol.","section":"Section III, Tables I and II; Section II.E; Section III.A"},{"comment":"The claim that GLoG-CSUnet adds only 15 parameters on Synapse and 30 on ACDC is not consistent with the architecture as described. In Figure 1(b), the outputs of the Gabor and LoG units are concatenated and then processed by 'multiple convolutional layers with GELU activation and layer normalization,' which introduces learnable convolution kernels and changes the input-channel dimension of the first convolutional layer. Counting only the scalar filter parameters (e.g., wavelength, orientation, sigma) omits these weights. Please report the total number of parameters and, if possible, FLOPs before and after the modification, and reconcile the count with Figure 1(b).","section":"Section II.A, Figure 1(b), Section IV"},{"comment":"The abstract states that GLoG-CSUnet achieves a 0.99% increase in Dice score for ACDC, but Table II reports 92.28% for the proposed method versus 91.37% for CS-Unet, which is a difference of 0.91 percentage points; for Synapse, the table values give 83.36 - 82.21 = 1.15 percentage points, not 1.14. The central quantitative claim must be internally consistent; please correct the abstract or the tables and specify the exact rounding convention.","section":"Abstract vs. Table II"}],"minor_comments":[{"comment":"The column header 'D' in Table I is undefined; it should be 'HD95' to match the metric described in Section II.D and the text in Section III.","section":"Table I"},{"comment":"The symbol λ is used for both the wavelength of the sinusoidal component and the spatial aspect ratio in the equation; the aspect ratio is conventionally denoted γ. Please clarify the notation.","section":"Section II.B.1, Eq. (1)"},{"comment":"The sentence fragment 'training dynamics.' appears in the experimental setup paragraph and should be removed or completed.","section":"Section II.E"},{"comment":"There are typos: 'sucsses' should be 'success' in Section IV, and 'a advancement' should be 'an advancement' in Section V.","section":"Section IV and V"},{"comment":"The related-work discussion would benefit from citing the original Swin Transformer paper alongside Swin-Unet, since the backbone is the Swin Transformer and the comparison is made against Swin-Unet.","section":"Section I.A"}],"recommendation":"major_revision","confidential_remarks":"The central idea is plausible and the paper is clearly written, but the empirical evidence needs strengthening: run-to-run variability must be quantified, the parameter-overhead claim must be reconciled with the architecture, and the abstract must match the tables. No concerns about citation loops or novelty disclosure; the reference list is appropriate for the scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper puts a few learnable Gabor and LoG filters into the patch embedding of a Swin-Unet variant and reports Dice gains of about 1 point on Synapse and ACDC. The idea is sensible and easy to test; the execution is clean enough to follow. But the reported gains are not backed by the kind of evidence needed to believe them: no error bars, no multiple seeds, and baselines are copied from earlier papers. As written, the central claim is plausible but not established.\n\nWhat's genuinely new: I don't know of a prior Swin-based segmentation model that inserts learnable Gabor/LoG units into the patch embedding. That specific integration is a reasonable extension of existing hybrid transformer-CNN work. The paper does a decent job of describing the architecture, gives the filter equations, runs ablations, and points to code.\n\nThe soft spots are mostly evidentiary. The 1.14-point Synapse gain and 0.91-point ACDC gain are smaller than typical seed-to-seed variation for these datasets, and the paper gives no way to assess that variation. The baselines are quoted from their original papers, so training protocols differ; the paper even says pretrained weights were used for baselines 'if provided,' while their model is trained from scratch. The ablation on ACDC shows a 0.45-point gap between Gabor-only and combined, which is the same order as the headline gain, so a single run per variant cannot establish that the combination is systematic.\n\nThe 'minimal overhead' claim is also under-specified. The learnable scalars are indeed 15 and 30, but concatenating filter outputs into the patch embedding increases the input channels of the following conv layers, adding thousands of parameters. That doesn't kill the idea, but '15 and 30 additional parameters' overstates the resource argument.\n\nThere are smaller issues: the abstract says 0.99% for ACDC while Table II says 0.91; the discussion has a typo ('D of 23.02') and 'succsses'; and the authors miss prior deep-learning work on learnable Gabor filters. These are fixable.\n\nWho benefits: anyone building transformer segmentation models on small medical datasets and wondering if radiomic-style inductive biases help. The paper is a reasonable starting point, not a finished statement.\n\nRecommendation: send it to peer review, not desk reject. Ask the authors to report multiple seeds with error bars, re-run at least the strongest baselines under the same protocol, reconcile the parameter count with the actual architecture, and fix the abstract/table mismatch.","headline":"A sensible but weakly supported incremental idea: learnable Gabor/LoG filters in a Swin-Unet patch embedding yield small Dice gains that need error bars and controlled baselines before they can be believed.","tokens_in":8251,"tokens_out":3275,"would_cite":false,"duration_ms":29701,"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":"Inserting 15 to 30 learnable Gabor and LoG filter parameters into a Swin-Unet patch embedding raises Dice scores on two medical segmentation benchmarks by 1.14 and 0.91 percentage points.","keywords":["medical image segmentation","vision transformer","learnable Gabor filter","Laplacian of Gaussian","radiomics","Swin-Unet","patch embedding","Dice Similarity Coefficient"],"falsifier":"Re-run GLoG-CSUnet and CSUnet with identical preprocessing, augmentation, optimizer settings, and training budget for at least ten random seeds on the same Synapse and ACDC splits; if the distribution of Dice scores overlaps so that the difference between the two models is smaller than one standard deviation, the claimed advantage is not distinguishable from training noise.","tokens_in":7159,"feed_emoji":"🩺","tokens_out":12070,"duration_ms":104538,"temperature":0.7,"pith_summary":"The paper tries to show that a Vision Transformer for medical image segmentation can be made better at local detail simply by inserting a small bank of learnable classical filters into its patch-embedding stage. GLoG-CSUnet does this with Gabor filters for texture and orientation and Laplacian-of-Gaussian filters for edges and boundaries, adapting them during training and concatenating their outputs with the normal convolutional patch embedding before the Swin-Transformer blocks. On the Synapse multi-organ CT benchmark it reports a mean Dice of 83.36%, which is 1.14 percentage points above the best comparison model, and on the ACDC cardiac MRI benchmark it reports 92.28%, a gain of 0.91 points over the same baseline. The reason this would matter is that the whole addition costs only 15 or 30 extra parameters, suggesting that radiomics-style local features can be injected into Transformers without the parameter overhead of a hybrid CNN-Transformer design.","feed_headline":"15 extra parameters add a one-point Dice gain in medical segmentation","feed_subtitle":"Adding learnable Gabor and LoG filters to a Swin-Unet embedding improves Synapse and ACDC results with negligible cost.","key_machinery":"The load-bearing mechanism is the learnable radiomic filter bank placed inside the Convolutional Patch Embedding. The Gabor layer applies 2D Gabor functions $G(x_\\theta,y_\\theta)=\\exp(-(x_\\theta^2+\\gamma^2 y_\\theta^2)/(2\\sigma^2))\\cos(2\\pi x_\\theta/\\lambda+\\psi)$ with trainable parameters $\\lambda,\\theta,\\psi,\\sigma,\\gamma$; the LoG layer applies $\\mathrm{LoG}(x,y)=-\\frac{1}{\\pi\\sigma^4}(1-\\frac{x^2+y^2}{2\\sigma^2})e^{-(x^2+y^2)/(2\\sigma^2)}$ with trainable $\\sigma$. These filter responses are concatenated with the patch embedding features and then passed through convolutional layers, GELU activation, and layer normalization before the Convolutional Swin Transformer blocks process the patches. The claim is that this small set of adaptive scalars supplies precisely the local texture, edge, and boundary information the global self-attention mechanism lacks.","core_discovery":"The central claim, on the paper's own terms, is that the local-information weakness of Vision Transformers in medical segmentation can be repaired at the embedding stage rather than by adding convolutional blocks throughout the network. In GLoG-CSUnet, the patch-embedding stage of the Convolutional Swin-Unet base is augmented with two adaptive filter units: a Gabor unit whose filters have learnable wavelength, orientation, phase, scale, and aspect ratio, and a LoG unit with a learnable scale, whose outputs are concatenated with the convolutional embedding before the transformer stages. With 2 Gabor and 5 LoG filters on Synapse and 5 of each on ACDC, the added trainable scalars number only 15 and 30; the reported mean Dice reaches 83.36% on Synapse and 92.28% on ACDC, with the largest per-organ gains on structures with fine or irregular boundaries such as the stomach and the right ventricle.","pith_inferences":["A test the paper leaves implicit: running GLoG-CSUnet and CSUnet several times with different random seeds on the same data splits would show whether the one-point Dice difference exceeds seed-to-seed variability; the paper reports single runs.","The learned filter parameters themselves are not analyzed; inspecting them after training could reveal whether the model discovers organ-specific orientations or scales, and whether those values transfer across datasets.","The same embedding filter bank could be evaluated on natural-image segmentation or detection tasks, where boundary and texture cues also matter, to see whether the benefit is specific to small medical datasets.","Because the paper only trains from scratch, a natural extension is to test the filter bank on top of a pre-trained ViT backbone; if the gain disappears, the method may be substituting for features that large-scale pre-training already provides."],"forward_implications":["If the reported gain is real, local feature extraction for Transformer segmentation can be added at the embedding stage rather than by deepening the convolutional component of a hybrid model.","With only 15 or 30 added scalars, the method would preserve the runtime and memory profile of the base model, making it usable in clinical settings with tight compute budgets.","The per-organ results suggest that the filter bank helps most on structures with weak or irregular boundaries, such as the stomach in CT and the right ventricle in cardiac MRI.","The paper's flexibility argument implies the same filter-bank insertion should transfer to other Transformer-based backbones, so the design is a plug-in rather than a full architecture rewrite."],"supporting_citations":[{"why":"Provides the Convolutional Swin-Unet base architecture that GLoG-CSUnet extends with its filter bank.","marker":"[12]"},{"why":"Supplies the Synapse multi-organ CT dataset and the standard 18/12 training/test split used for evaluation.","marker":"[17]"},{"why":"Supplies the ACDC cardiac MRI dataset with its 70/10/20 train/validation/test protocol and three cardiac structures.","marker":"[18]"},{"why":"Defines the Swin-Unet Transformer backbone that the base model builds on and contributes baseline numbers in both comparison tables.","marker":"[10]"},{"why":"Introduces the TransUNet baseline and provides the common Synapse protocol and baseline scores quoted for comparison.","marker":"[8]"},{"why":"Contributes the MT-UNet baseline and its mixed-attention design, which the paper compares against in Tables I and II.","marker":"[11]"},{"why":"Defines the 2D Gabor function used to construct the learnable Gabor filter layer.","marker":"[16]"},{"why":"Motivates the use of standardized convolutional filtering for radiomics, grounding the choice of LoG and Gabor features.","marker":"[14]"}],"fun_headline_variants":["15 extra parameters add 1.14% Dice gain in Synapse","Gabor+LoG filters at embedding add 1% Dice with 15 params","Embedding-stage filters fix ViT local gap: 1% Dice gain","Tiny learnable filters refresh Swin-Unet embeddings for segmentation","15 params for a point of Dice: Gabor/LoG embedding trick"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a reported improvement of about one percentage point in Dice is larger than the ordinary run-to-run variation of these models, since the comparison pits a single run of the new model against baseline numbers quoted from earlier papers without error bars.","fun_headline_variants_meta":{"raw":{"variants":["15 extra parameters add 1.14% Dice gain in Synapse","Gabor+LoG filters at embedding add 1% Dice with 15 params","Embedding-stage filters fix ViT local gap: 1% Dice gain","Tiny learnable filters refresh Swin-Unet embeddings for segmentation","15 params for a point of Dice: Gabor/LoG embedding trick"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":3064,"prompt_tokens":1024,"completion_tokens":2040,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":1940}},"tokens_in":640,"tokens_out":2040,"duration_ms":13682,"temperature":1.0,"reasoning_tokens":1940,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:04:26.152804+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run GLoG-CSUnet and CSUnet with identical preprocessing, augmentation, optimizer settings, and training budget for at least ten random seeds on the same Synapse and ACDC splits; if the distribution of Dice scores overlaps so that the difference between the two models is smaller than one standard deviation, the claimed advantage is not distinguishable from training noise.","supporting_citations":[{"cited_title":"Optimizing vision trans- formers for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the Convolutional Swin-Unet base architecture that GLoG-CSUnet extends with its filter bank."},{"cited_title":"Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge,","cited_arxiv_id":null,"evidence_quote":"Supplies the Synapse multi-organ CT dataset and the standard 18/12 training/test split used for evaluation."},{"cited_title":"Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?,","cited_arxiv_id":null,"evidence_quote":"Supplies the ACDC cardiac MRI dataset with its 70/10/20 train/validation/test protocol and three cardiac structures."},{"cited_title":"Mixed transformer u-net for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Contributes the MT-UNet baseline and its mixed-attention design, which the paper compares against in Tables I and II."},{"cited_title":"In search of a general picture processing operator,","cited_arxiv_id":null,"evidence_quote":"Defines the 2D Gabor function used to construct the learnable Gabor filter layer."},{"cited_title":"Standardised convolutional filtering for radiomics","cited_arxiv_id":"2006.05470","evidence_quote":"Motivates the use of standardized convolutional filtering for radiomics, grounding the choice of LoG and Gabor features."}],"review_version":1}